Integrating feature selection and explainable CNN for identification and classification of pests and beneficial insects

Abstract

Reliable identification of agricultural pests and beneficial insects is crucial for sustainable crop protection and ecological balance, yet most vision-based models remain black boxes and require high-dimensional features. This paper proposes an explainable hybrid insect-classification framework that combines convolutional neural network (CNN) feature extraction with a dual–XAI feature selection strategy. SHapley Additive exPlanations (SHAP) and Permutation Feature Importance (PFI) are applied in parallel to rank handcrafted and CNN-derived features, and their intersection yields a compact, biologically meaningful subset for final classification. The selected features are evaluated using lightweight classifiers and a hybrid ensemble, enabling accurate inference under field variability. Experiments on a curated, balanced dataset of four classes (Colorado potato beetle, green peach aphid, seven-spot ladybird, and healthy leaves) collected under diverse lighting and background conditions achieve 96.7% overall accuracy, with precision, recall, and F1-scores all above 96%. Importantly, performance remains stable when reducing dimensionality, retaining $\ge$90% accuracy using only the top 11 hybrid-selected features. These results demonstrate that integrating SHAP and PFI improves both robustness and interpretability, supporting practical deployment for automated pest monitoring and precision agriculture.

Introduction

Accurate and timely identification of insect pests and beneficial insects plays a vital role in integrated pest management (IPM) and sustainable agriculture. Insect pests, if not immediately identified and controlled, can have devastating impacts on crop yields and food security  ^1,2. In contrast, beneficial insects such as pollinators and natural predators contribute significantly to ecosystem balance and crop productivity. Misclassification or late detection can result in misuse of agrochemicals, increased production costs, ecological imbalance, and decreased agricultural production  ^3,4. The Timely and accurate identification of insect pests and beneficial species is critical to implementing effective Integrated Pest Management (IPM) strategies in sustainable agriculture  ^1,5. Incorrect identification can result in either excessive pesticide application, harming pollinators and natural predators, or insufficient pest control, leading to severe crop losses. According to the Food and Agriculture Organization  ^6,7, global economic losses due to plant pests and diseases exceed $220 billion annually, with insect pests alone accounting for more than 31.8% of this figure. Among the most damaging insect pests in arable crops are the Colorado potato beetle CPB, (Leptinotarsa decemlineata) and the green peach aphid Myzus persicae. CPB larvae are known to completely defoliate potato plants, potentially reducing yields by more than 50% if not effectively managed^8,9. Moreover, aphids such as M. persicae are efficient vectors of viral diseases like Potato Virus Y (PVY), with annual economic losses in Europe alone estimated at over €180 million^10,11.

Traditionally, pest monitoring has relied on manual scouting and expert identification, followed by widespread use of synthetic pesticides. Although still common practice, this approach is not only labor-intensive and time-consuming but also contributes significantly to environmental degradation, pesticide resistance, and non-target species decline^1,12. There is therefore a growing need for automated, intelligent detection systems capable of distinguishing between pest and beneficial insects efficiently and accurately. Recent developments in computer vision, sensor technologies, and machine learning (ML) have enabled the design of automated insect detection systems. Early systems relied on basic image processing methods–such as thresholding, morphological operations, and edge detection–to isolate and classify insect features^13,14. For instance,¹ introduced a multifractal-based segmentation technique for whiteflies, and^9,15 employed Mahalanobis distance-based classification for pests on sticky traps. However, these methods are often unreliable under field conditions due to varying lighting, cluttered backgrounds, and occlusions, which limit their robustness and scalability^4,16. Recent advancements in computer vision have significantly enhanced the ability to detect and classify insects in agricultural settings. Convolutional neural networks (CNNs) and deep learning architectures have demonstrated high accuracy in identifying and localizing insect pests across diverse crop types^17,18. For instance, transformer-based models have shown superior performance in segmenting small insect regions within complex backgrounds^2,19. Hybrid models combining CNNs with attention mechanisms, such as Vision Transformers (ViTs), have further improved classification accuracy and robustness²⁰. Moreover, domain adaptation techniques have been successfully applied to bridge the gap between laboratory-controlled and real-world field images¹⁵. Researchers have also explored multi-modal fusion, integrating RGB and hyperspectral images to boost insect detection performance under varying environmental conditions²¹. Despite these advancements, challenges persist in segmenting small and occluded insects, motivating the exploration of ensemble and hybrid models that combine multiple feature types and classifiers⁶. These studies collectively highlight the growing promise of computer vision-based insect detection as a critical tool for precision pest management in sustainable agriculture.

Image processing forms the foundational stage in deep learning workflows, playing a crucial role in extracting informative visual features for downstream tasks such as object detection and classification. Recent advances in computer vision, including transformer-based architectures^2,20, have further highlighted the importance of robust preprocessing techniques in ensuring that deep models can effectively learn representations from diverse and complex visual inputs. As agricultural and ecological datasets often present high variability in background, lighting, and object scales, effective image processing ensures the reliability of learned features across different conditions¹⁷. Typical preprocessing steps in deep learning pipelines include resizing images to standard dimensions, normalization to maintain consistent intensity ranges, and data augmentation to improve generalization^18,22. Techniques such as histogram equalization and adaptive thresholding enhance contrast, while color space transformations (e.g., RGB to HSV) enable models to leverage color information for more accurate segmentation. For insect detection, these preprocessing steps ensure that small, low-contrast insect bodies are visible and distinguishable against complex plant backgrounds. Effective image processing directly impacts the semantic segmentation accuracy of deep learning models, particularly when dealing with small or occluded insect objects⁶. Models trained on preprocessed images achieve significantly higher precision and recall, as preprocessing reduces noise and highlights relevant regions of interest. Moreover, data augmentation techniques such as rotations, translations, and brightness adjustments expand the effective dataset size, helping the model to generalize better and avoid overfitting to specific capture conditions²¹. Given the complexity of real-world agricultural environments, dataset-specific preprocessing strategies are essential to maximize performance in deployment scenarios²⁰. Factors such as varying illumination, crop type differences, and natural occlusions necessitate tailored preprocessing pipelines that adapt to the specific challenges of each dataset. Future work in insect detection should focus on robust preprocessing methods that consider these variations, bridging the gap between laboratory-controlled datasets and dynamic field conditions². Such efforts are key to developing scalable, reliable pest-detection systems for precision agriculture.

To improve robustness, traditional ML techniques have been widely applied in pest detection systems. These models generally use manually extracted features such as shape descriptors, color histograms, texture (e.g., GLCM), and Histogram of Oriented Gradients (HOG) to represent the visual properties of insects. Algorithms such as Support Vector Machines (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), and Naïve Bayes (NB) have demonstrated competitive performance, particularly in resource-constrained environments where lightweight computation is necessary⁴. For example,⁸ reported over 86% accuracy in aphid detection using HOG features with SVM, while⁵ achieved more than 91% accuracy using various classical ML classifiers trained on shape features. However, these classical approaches face challenges due to the presence of high-dimensional and often redundant features. Including irrelevant features can increase computational load, slow down model inference, and impair generalization performance^23,24. Additionally, no universal feature set exists that performs optimally across all insect categories and conditions. Thus, effective feature selection is essential for maximizing classification accuracy while minimizing complexity and resource use. Although deep learning–particularly convolutional neural networks (CNNs)–has become dominant in image classification tasks due to its ability to automatically learn features, such models require extensive labeled data and computational resources²⁵. Their deployment on low-cost edge devices remains limited, especially in agricultural contexts with constrained infrastructure^14,26. As a result, further exploration of classical ML models augmented by intelligent feature selection is warranted. Recent advancements in machine learning (ML) have significantly improved the detection and classification of insect species compared to traditional thresholding and rule-based image analysis methods^15,27. Classical ML models, such as support vector machines (SVMs) and artificial neural networks (ANNs), have leveraged handcrafted features–such as shape descriptors, texture features, and color histograms–to distinguish insect classes in agricultural images¹⁸. For example, studies using SVM models with histogram of oriented gradients (HOG) and morphological descriptors have achieved classification accuracies above 85% for species like the Colorado potato beetle and green peach aphid²¹. More advanced ML strategies employ ensemble models and multi-layer classifiers, which integrate multiple base classifiers to enhance accuracy and robustness²⁸. While these approaches rely heavily on precise feature extraction and preprocessing, deep learning models have shifted the focus toward automated feature learning from raw pixel data. Convolutional neural networks (CNNs), in particular, outperform classical methods by simultaneously learning both region-of-interest (ROI) segmentation and classification tasks²⁹. This profound learning transition offers promising avenues for real-time pest monitoring in precision agriculture.

Insect classification plays a pivotal role in sustainable agriculture, particularly in monitoring harmful pests such as the Colorado Potato Beetle (CPB) and aphids, and preserving beneficial species like ladybirds ^21,30. Effective identification of these species is crucial for integrated pest management (IPM) strategies and reducing overreliance on pesticides ². Although machine learning (ML) techniques have been increasingly employed for insect detection ^1,4, most existing approaches suffer from two key limitations: (1) reliance on high-dimensional, non-informative features, and (2) the use of black-box models lacking interpretability ^5,12, which impedes trust and real-world deployment. To address these challenges, we propose a novel hybrid feature selection strategy that leverages the strengths of two explainable AI (XAI) tools–SHapley Additive exPlanations (SHAP) ^31,32 and Permutation Feature Importance (PFI) ⁴. Our approach systematically identifies the most relevant and non-redundant features, enhancing both classification performance and interpretability. The selected feature subsets are evaluated using classical ML classifiers such as Support Vector Machines (SVM), Random Forests (RF), k-Nearest Neighbors (KNN), and Naïve Bayes (NB), enabling low-latency inference suitable for resource-constrained field conditions ¹⁵. This work contributes not only a new hybrid-XAI methodology but also a practical solution toward deployable smart agricultural monitoring systems. To our knowledge, no prior work has fused SHAP and PFI for agricultural insect detection, nor has such an approach been assessed for its trade-offs in accuracy and efficiency on edge devices ^23,33.

Research gap: Recent studies have achieved notable progress in agricultural insect recognition through (i) classical ML pipelines that combine handcrafted features (HOG, texture, color, shape) with lightweight classifiers, and (ii) deep CNN/transformer-based models that deliver high accuracy on controlled or benchmark datasets^1,2,5. A few works have also begun to use XAI tools such as SHAP or LIME to provide post-hoc visual explanations of trained models^31,32. Nevertheless, several critical challenges remain insufficiently addressed. First, most existing systems emphasize predictive accuracy while relying on high-dimensional or redundant feature spaces, and they rarely justify why particular features are selected; this limits transparency, increases computational cost, and weakens generalization under field variability. Second, robustness across real-world conditions–such as changing illumination, cluttered multi-crop backgrounds, partial occlusions, and camera-device diversity–remains a major bottleneck, particularly for models trained on narrow or laboratory-biased datasets^4,13,33. Third, the literature is heavily skewed toward pest-only detection, with limited attention to beneficial insects and multi-species ecological interactions, which are essential for sustainable IPM and avoiding unintended harm to pollinators or predators^8,34. Fourth, while XAI is increasingly used for interpretation, it is seldom integrated into the feature selection stage; to the best of our knowledge, no prior insect-classification work has fused complementary explainability tools (e.g., SHAP and PFI) into a hybrid, high-confidence selection mechanism, leaving the synergistic potential of dual-XAI selection unexplored. Finally, the field still lacks compact, interpretable, and edge-ready models that can maintain stable accuracy with minimal features and low latency on resource-constrained hardware. Collectively, these gaps underscore the need for robust, explainable, and computationally efficient insect classification frameworks that can identify both harmful and beneficial species across diverse field conditions.

Research objectives:This study aims to develop and evaluate a hybrid feature selection strategy that leverages two prominent explainable AI (XAI) tools–Permutation Feature Importance (PFI) and SHapley Additive exPlanations (SHAP)–to identify the most relevant and non-redundant features for effective insect classification. The objective is to enhance the interpretability and performance of classical machine learning classifiers, such as Support Vector Machines (SVM), Random Forests (RF), k-Nearest Neighbors (KNN), and Naïve Bayes (NB), particularly in distinguishing between harmful insects (e.g., Colorado potato beetles and aphids) and beneficial species (e.g., ladybirds). By systematically evaluating different feature subsets, this research investigates the impact of XAI-guided selection on classification accuracy and computational efficiency. Furthermore, the study aims to design a lightweight and efficient detection system suitable for deployment on low-cost hardware in real-world agricultural environments.

Research contributions:The main contributions of this study are summarized as follows:

• 1. We introduce a dual-XAI hybrid feature selection strategy that integrates SHAP and Permutation Feature Importance (PFI) in parallel and selects features via their intersection, yielding a compact and highly reliable subset. To the best of our knowledge, this SHAP+PFI fusion has not been previously explored for agricultural insect classification.

• 2. We demonstrate that the proposed dual-XAI selection improves both accuracy and stability under aggressive dimensionality reduction, retaining $\ge 90\%$ accuracy using only the top-11 features, thereby validating robustness beyond conventional single-method selection.

• 3. We benchmark the hybrid-XAI pipeline against standard filter, wrapper, and embedded methods across four lightweight classifiers (SVM, RF, KNN, NB) and a hybrid ensemble, providing a comprehensive trade-off analysis between interpretability, performance, and computational cost.

• 4. We validate the framework on a real farm dataset containing both harmful pests (CPB, aphids) and a beneficial insect (ladybird), collected under diverse field lighting and background conditions, addressing a key realism gap in prior studies.

• 5. We provide feature-level explanations aligned with entomological traits, enabling transparent decision support suitable for low-cost edge deployment in precision agriculture and integrated pest management.

This paper is organized as follows: Section 2 reviews relevant literature on insect detection, machine learning, and feature selection. Section 3 describes the dataset, feature extraction process, and the proposed XAI-based feature selection framework. Section 4 presents the experimental results and performance comparisons. Section 5 discusses the findings and their practical implications. Finally, Section 6 concludes the paper and outlines directions for future research.

Related work

Recent progress in agricultural insect detection spans four main directions: (i) image-based detection using handcrafted features and classical machine learning, (ii) deep learning models for end-to-end recognition, (iii) feature selection and dimensionality reduction to improve efficiency, and (iv) explainable AI (XAI) for transparency. This section summarizes what has already been achieved in each direction and clarifies the remaining limitations that motivate our hybrid-XAI framework.

Early automated insect recognition systems relied on conventional image processing and handcrafted descriptors. Techniques such as color thresholding, edge detection, and region segmentation were commonly used to isolate insects from foliage or sticky traps^8,12. These pipelines were later coupled with low-cost sensing platforms for greenhouse or field monitoring^13,35. Classical ML classifiers–such as Support Vector Machines (SVM), Random Forests (RF), K-Nearest Neighbors (KNN), Naïve Bayes (NB), and Gradient Boosting Machines (GBM)–have shown reliable performance when driven by descriptive features (HOG, GLCM, LBP, color moments, and shape metrics)^5,30. For example, HOG + SVM pipelines have achieved strong aphid classification results, while RF-based models remain robust to moderate noise and feature interactions. These achievements demonstrate that classical ML remains effective for lightweight insect recognition, particularly under constrained computation.

Deep learning has recently become dominant due to its ability to learn hierarchical representations directly from raw images. CNN architectures and transfer learning using pretrained backbones (e.g., VGG, Inception, ResNet) have significantly improved insect classification accuracy, especially on large curated datasets^33,36. Vision Transformer (ViT) and attention-augmented CNN variants further enhance detection performance under cluttered backgrounds and for small objects^2,20. These works establish that deep models achieve state-of-the-art accuracy in controlled settings and can detect subtle morphological differences across insect taxa. However, despite these achievements, deep models often demand large labeled datasets and high computational resources, which limit deployment on low-cost edge devices and in smallholder farming contexts^3,15. Moreover, their black-box nature remains a major barrier to trust and adoption in safety-critical decision workflows such as integrated pest management.

To improve efficiency and avoid redundancy, many studies incorporate feature selection before training classical ML classifiers. Filter methods (e.g., mutual information, chi-square, and correlation ranking), wrapper techniques (e.g., recursive feature elimination), and embedded approaches (e.g., LASSO and tree-based importance) have been widely adopted^13,23. These approaches reduce dimensionality and often increase accuracy by removing irrelevant descriptors. Prior results confirm that compact feature sets can accelerate inference and reduce overfitting, which is essential for real-time monitoring applications. Nevertheless, most selection strategies remain statistical or model-driven, without offering a biologically interpretable justification for why a feature matters. As a result, selected features may still be unstable across environments, and end users cannot easily verify whether retained cues align with insect morphology or ecology.

XAI has emerged as a critical tool for opening black-box models and supporting trustworthy agricultural decision-making. SHAP assigns game-theoretic contribution values to features for both local and global explanations, while Permutation Feature Importance (PFI) quantifies how prediction performance changes when a feature is permuted^31,32. Several agricultural studies have used SHAP or LIME to interpret crop or disease models^1,37, and a small number of insect studies have introduced XAI mainly as post-hoc visualization. However, two limitations remain clear in the current literature. First, XAI is rarely integrated into the feature selection stage; it is typically applied only after training to explain outcomes rather than to guide feature reduction. Second, existing works almost exclusively use a single XAI method, leaving the complementary strengths of multiple explainability tools unexplored. To the best of our knowledge, no prior insect classification study has combined SHAP and PFI into a hybrid, intersection-based selection pipeline that explicitly prioritizes stable, high-confidence features.

In summary, prior studies have achieved reliable insect recognition using handcrafted features and classical ML, and high accuracy using CNN/transformer models with transfer learning. Feature selection has improved efficiency but remains largely non-interpretable, and XAI has improved transparency but is mostly applied post-hoc and in isolation. The remaining open challenges include (i) reducing redundancy while preserving biological interpretability, (ii) ensuring stability under real-field variability, (iii) supporting lightweight edge deployment, and (iv) jointly addressing pests and beneficial insects. Our work addresses these gaps by integrating SHAP and PFI directly into hybrid feature selection, yielding a compact, explainable subset that enables robust, field-deployable insect classification.

Methodology

This section presents our hybrid explainable AI (XAI)-based feature selection framework, designed to improve insect classification by identifying informative and non-redundant features. The methodology consists of four main stages: data preprocessing, feature extraction, hybrid feature selection using SHAP and PFI, and model evaluation.

Dataset

This study utilizes a curated dataset comprising labeled images of pest and beneficial insects on key crops. Specifically, we focused on two economically essential pest species–the Colorado potato beetle (CPB, Leptinotarsa decemlineata) and the green peach aphid (Myzus persicae)–and one beneficial species, the seven-spot ladybird (Coccinella septempunctata). These insects were imaged in association with common crops: potato (Solanum tuberosum), eggplant (Solanum melongena), tomato (Solanum lycopersicum), gourd (Beta vulgaris), chili pepper (Capsicum annuum) and bitter melon (Luffa cylindrica syn. L. aegyptiaca). All images are resized to a uniform resolution and augmented using horizontal flipping, rotation, and brightness adjustment to improve generalizability. Features are extracted using pre-trained CNN backbones (e.g., MobileNetV2), then flattened and normalized.

Data collection

The crop plants were cultivated in open field plots, and images were collected during 2023–2025 across experimental plots located in Talshar, Mohespur, Jhenidah, Bangladesh (23.5451$^{\circ }$N, 89.1683$^{\circ }$E). These experimental farms are privately owned and managed by the authors, Dr. Tawfikur Rahman and Dr. Nibedita. Therefore, no institutional or governmental permissions were required for the collection of plant or insect specimens, as all research activities were conducted within privately owned research fields under the authors’ supervision. Insect specimens were collected from three nearby experimental farms situated in the central, northeastern, and southwestern regions of Talshar, Mohespur, Jhenidah, encompassing diverse agroecological zones to capture real-world variability. Images were captured using multiple devices, including Vivo 30 Pro Max, Vivo 17 Pro Max, Canon DS126291, and Samsung 16 Pro, to ensure high-quality imaging across different sensor types. For this study, six major vegetable crops commonly grown in Bangladesh–potato (Solanum tuberosum), pointed gourd (Trichosanthes dioica), eggplant (Solanum melongena), bitter melon (Momordica charantia), tomato (Solanum lycopersicum), and chili pepper (Capsicum annuum)–were selected as host plants for insect imaging. These crops are vital to smallholder farmers and are known to attract a wide variety of pest and beneficial insect species. Four developmental stages–egg, larva, pupa, and adult–of both the Colorado potato beetle (CPB, Leptinotarsa decemlineata) and the seven-spot ladybird (Coccinella septempunctata) were collected, along with adult green peach aphids (Myzus persicae). Each insect species was housed in separate ventilated rearing cages (36 $\times$ 36 $\times$ 6 cm, 500 g), constructed from fine mesh to ensure airflow and secure containment. Insects were fed their respective host plants: CPBs on eggplant, aphids on all four crop types, and ladybirds with live aphids as prey. Crop plants used for insect feeding and imaging were grown in soil-filled plastic pots and regularly irrigated to maintain healthy foliage and natural plant-insect interactions. To promote authentic insect behavior during imaging, greenhouse conditions mimicked local climate variations, maintaining temperatures between 25$^{\circ }$C and 40$^{\circ }$C, typical of summer and pre-monsoon conditions in Bangladesh. Insects were kept for 1–4 weeks before being replaced with freshly collected specimens to ensure ongoing diversity and active behavior. This generated a dynamic dataset that realistically represents field-level insect–plant interactions. To enhance robustness, images were collected under diverse real-world conditions, including varying natural illumination, shadows, and complex crop backgrounds in open-field settings. Data acquisition was performed at different times of day (morning, afternoon, and evening) and across multiple host plants and locations to capture realistic background clutter and occlusions.

Imaging setup

To create a dataset that accurately captures the diversity and complexity of real-world agricultural conditions in Bangladesh, insect images were collected in directly in the field across multiple locations. In the farm-based laboratory, a custom-built imaging tent (124 $\times$ 124 $\times$ 124 cm) was assembled using black Mylar fabric to eliminate external light interference and reflections. A modular LED lighting system (LE1200 series) was integrated within this enclosure, providing versatile illumination with LEDs emitting at 450 nm, 470 nm, 660 nm, and 730 nm, along with 7000K white light to replicate daylight conditions. Lighting intensities were carefully adjusted between 500 and 1500 lux, as measured by a calibrated lux meter (LX1330B), ensuring consistent lighting across all image captures. Crop plants including pointed gourd, eggplant, bitter melon, and capsicum, each infested with aphids, were placed inside the enclosure for imaging. Individual CPBs and seven-spot ladybirds were manually positioned on the plants to replicate natural infestation scenarios: CPBs, which target solanaceous crops, were imaged exclusively on eggplant, while aphids and ladybirds were imaged on all four crop types. A range of camera angles and distances was employed to ensure dataset variability, with close-up shots capturing intricate morphological features such as leg articulation, body segmentation, and wing patterns.

Sample images from the farm field, including various poses and workflows of beneficial and pest insects, are shown in Fig. 1. These images reflect real infestation scenarios, varied lighting conditions, and various crop-insect interactions, forming a comprehensive dataset suitable for robust insect classification and detection tasks. This imaging workflow was designed to mirror local surveillance practices in Bangladesh, where pest scouting is typically performed visually during daylight hours, and control decisions are based on economic threshold levels advised by agricultural officers. Our setup emulates this process digitally, enabling automation through machine learning while maintaining biological and ecological realism.

Fig. 1

The various pest and beneficial insect poses.

In total, approximately 500 CPBs, 500 ladybirds, and 500 aphid colonies were imaged under various lighting intensities and positions. To enhance dataset variability further, imaging was conducted at three distinct times of day–morning (8–10 a.m.), afternoon (1–3 p.m.), and evening (4–6 p.m.)–to reflect typical field scouting hours and natural lighting changes. Field images were sourced from three agricultural surveillance sites: Jhenidah, Chuadanga, and Jessore, representing key vegetable-growing regions with differing pest dynamics. Local extension officers and farmers provided insights into typical infestation patterns and validated pest presence, ensuring the dataset’s practical relevance. Field images included both insect-infested and uninfested leaves to provide negative examples crucial for model learning. A combination of two smartphones (VIVO 30pro and VIVO17pro) and one digital camera (Canon EOS 2000D) was used for image acquisition, each capturing approximately 1000 images. All images were taken at a native resolution of 6000 $\times$ 4000 pixels, but subsequently resized to 800 $\times$ 500 pixels during preprocessing to optimize computational efficiency while retaining insect detail for robust machine learning model training.

Dataset composition

The complete dataset consists of 1500 labeled images collected directly from real farm environments across central Bangladesh, ensuring strong field relevance and natural variability. It covers four classes: two pest categories–Colorado potato beetle (CPB, Leptinotarsa decemlineata) and green peach aphid (Myzus persicae)–one beneficial insect category–seven-spot ladybird (Coccinella septempunctata)–and a healthy, insect-free leaf class serving as a negative control. In the full dataset, approximately 500 images were collected for each insect class (CPB, ladybird, aphid), captured across multiple host crops and imaging conditions, while healthy-leaf images were included to support reliable discrimination between infested and uninfested plants. For experimental evaluation and fair classifier comparison, we constructed a balanced benchmark subset of 1000 images by selecting 250 samples per class (total 250 CPB, 250 ladybird, 250 aphid, and 250 healthy leaves). Ladybird and aphid samples were evenly distributed across pointed gourd, bitter melon, and capsicum (approximately 83 images per crop per class in the benchmark subset), ensuring crop-level balance and reducing host-plant bias. Including healthy, insect-free leaves is essential for practical deployment, as it enables the model to distinguish true infestations from visually similar non-infested foliage under real Bangladeshi farm conditions.

Preprocessing and annotation

All original images, captured at 6000 $\times$ 4000 pixels with DSLR cameras, were downscaled to 800 $\times$ 500 pixels to reduce memory usage and accelerate model training. To eliminate irrelevant visual information and focus on the insect-plant regions, a manual crop-and-select technique was employed, adapted from⁴. In this method, each image was segmented into four quadrants, and only the sections that contained both insects and crop foliage were retained. This approach ensured the exclusion of noisy backgrounds such as soil, shadows, or unrelated farm equipment. Annotation was performed using the open-source tool LabelImg, which enabled accurate labeling of each insect instance through rectangular bounding boxes. Annotations were stored in XML format following the Pascal VOC standard, ensuring compatibility with widely used machine learning libraries. This meticulous preprocessing pipeline helped ensure clean, consistent data inputs critical for robust insect detection and classification under diverse Bangladeshi farm conditions.

Feature extraction

A comprehensive set of features was extracted from the annotated images to enable accurate classification of insect species. These features were carefully selected to capture visual and morphological characteristics relevant to insect identification. Histogram of Oriented Gradients (HOG) was used to extract edge and shape information, allowing the model to distinguish between different outlines of the insect body and structural characteristics, Color histograms were used to capture variations in color patterns across species, providing another key differentiator, particularly useful in distinguishing insects like the seven-spot ladybird from aphids. Additionally, shape descriptors, including area, perimeter, aspect ratio, and roundness, were computed to quantify each insect’s morphological traits. These shape-based metrics are beneficial for separating life stages (e.g., larva vs. adult) and identifying insects with similar color profiles. The diversity and richness of this feature set, derived from a dataset collected under both controlled and field conditions, significantly enhance the generalizability and performance of machine learning models for insect detection and classification.

Figure 2 represents various feature selection techniques were applied to the extracted feature set to enhance model performance, reduce dimensionality, and improve interpretability. These methods were categorized into three main types: filter, wrapper, and embedded approaches. Filter methods were used to independently evaluate the relevance of each feature concerning the target class, without involving any specific machine learning algorithm. In particular, correlation-based feature selection (CFS) was employed to identify features with a high correlation with the class label but low intercorrelation among themselves, thus minimizing redundancy. Mutual information, another filter-based method, quantified the statistical dependence between individual features and the class labels, helping to capture non-linear relationships.

Fig. 2

The various feature selection techniques.

In addition to filter methods, wrapper methods were utilized to assess feature subsets by repeatedly training and evaluating a model. Among these, Recursive Feature Elimination (RFE) was adopted, which recursively removes the least important features and re-evaluates model performance. RFE is particularly effective at identifying the most informative feature combination, but it can be computationally intensive. Finally, embedded methods integrate feature selection into the model training process. Lasso Regression (Least Absolute Shrinkage and Selection Operator) was used to enforce sparsity by penalizing the absolute size of coefficients, effectively eliminating less critical features. Additionally, feature importance scores derived from the Random Forest algorithm were employed to rank and select features based on their contribution to reducing classification error across decision trees. These embedded methods offer the advantage of balancing performance with interpretability. By combining these diverse techniques, the most relevant and non-redundant features were selected, ensuring that the final model is accurate, efficient, and generalizable across different insect species and imaging conditions.

Feature selection plays a crucial role in enhancing the accuracy and efficiency of machine learning models, especially for image-based insect classification. This section outlines three significant categories of feature selection techniques: filter, wrapper, and embedded methods. Mutual Information quantifies the statistical dependency between a feature and the target class. It is defined as in Eq. (1).

$$\begin{aligned} MI(X; Y) = \sum _{x \in X} \sum _{y \in Y} P(x, y) \cdot \log \left( \frac{P(x, y)}{P(x) \cdot P(y)} \right) \end{aligned}$$

(1)

Where $P(x, y)$ is the joint probability of feature $X$ and class $Y$, and $P(x)$, $P(y)$ are the marginal probabilities. Higher MI scores indicate stronger relevance to the target.

Recursive Feature Elimination (RFE) iteratively trains a machine learning model and removes the least significant feature based on weight coefficients or impurity scores. Recursive Feature Elimination (RFE) is a wrapper-based feature selection method that removes the least important features based on the model’s coefficients or feature importance scores. Below is a mathematical interpretation of the RFE process defined in Eqs. (2) and (3). Let the dataset be represented by a feature matrix $X \in \mathbb {R}^{n \times p}$, where $n$ is the number of samples and $p$ is the number of features. The target vector is $y \in \mathbb {R}^n$.

$$\begin{aligned} \beta = [\beta _1, \beta _2, \ldots , \beta _p]^T \end{aligned}$$

(2)

Denote the coefficients of a linear model trained on $X$. The model is trained using ordinary least squares (OLS):

$$\begin{aligned} \hat{\beta } = \arg \min _{\beta } \Vert y - X\beta \Vert _2^2 \end{aligned}$$

(3)

The RFE process involves the following iterative steps. Train a linear model on the current feature set. Compute importance scores using the squared magnitude of coefficients defined in Eq. (4).

$$\begin{aligned} I_j = \beta _j^2 \quad \text {for } j = 1, 2, \ldots , p \end{aligned}$$

(4)

Identify the feature with the smallest score defined in Eq. (5).

$$\begin{aligned} j^* = \arg \min _j I_j \end{aligned}$$

(5)

Remove feature $j^*$ from $X$ and update $\beta$. Let the new reduced feature set be $X' \in \mathbb {R}^{n \times (p-1)}$. Repeat the process until $p'$ desired features remain. RFE systematically reduces dimensionality by iteratively pruning the least-informative features using a mathematical criterion based on model weights.

Lasso regression integrates feature selection during model training using L1 regularization. The objective function is defined as in Eq. (6).

$$\begin{aligned} \min _{\beta } \left( \frac{1}{2n} \sum _{i=1}^n (y_i - X_i \cdot \beta )^2 + \lambda \sum _{j=1}^p |\beta _j| \right) \end{aligned}$$

(6)

Where $\beta$ are model coefficients, and $\lambda$ controls the strength of regularization. Coefficients shrunk to zero are considered unimportant and thus removed. Random Forests determine feature importance based on the total decrease in impurity (e.g., Gini index) contributed by each feature, as defined in Eq. (7).

$$\begin{aligned} G = 1 - \sum _{k=1}^K p_k^2 \end{aligned}$$

(7)

Where $p_k$ is the proportion of class $k$ samples at a node? Features that contribute more to impurity reduction across trees are considered more important. These feature selection methods enable the identification of the most informative attributes from image data, enhancing model generalization and performance. Mutual Information is fast and model-independent; RFE is model-specific but often yields high performance; and embedded methods like Lasso and Random Forest combine training and selection into one efficient step.

Hybrid feature selection pipeline

The proposed pipeline combines SHapley Additive exPlanations (SHAP) and Permutation Feature Importance (PFI) to select features that are both relevant and non-redundant. This hybrid approach addresses the limitations of traditional feature selection methods by introducing model interpretability into the selection process.

Start: Input dataset-$X, y$

Step 1: Train base classifier -$f$

Step 2a: Compute SHAP values $\rightarrow$ rank features $\rightarrow$ select top-$k_1$ $\rightarrow$ set $F_{\text {SHAP}}$

Step 2b: Compute PFI scores $\rightarrow$ rank features $\rightarrow$ select top-$k_2$ $\rightarrow$ set $F_{\text {PFI}}$

Step 3: Intersect $F_{\text {SHAP}} \cap F_{\text {PFI}} = F_{\text {hybrid}}$

Step 4: Final reduced feature matrix $X_{\text {selected}}$

Output: Features ready for classification

To enhance insect detection across varied crop types, a hybrid classification model was developed by combining three distinct machine learning algorithms: SVM, RF, and GBM. Each algorithm processes the same input features and independently outputs a classification decision. Figure 3 illustrates the algorithm of a hybrid machine learning model designed for insect detection across various crop types. Before classification, feature selection techniques such as mutual information and Lasso regression are applied to reduce dimensionality and retain only the most informative features.

Fig. 3

The algorithm of a hybrid machine learning model designed for insect detection across varied crop types.

The predictions of the three classifiers are aggregated using a majority voting ensemble method, where the final class label is the one predicted most frequently by the models. This ensemble approach exploits the complementary strengths of each algorithm–SVM’s margin optimization, RF’s robustness to noise and overfitting, and GBM’s ability to model complex patterns–thus improving overall detection accuracy and generalization. This design is particularly effective for datasets involving multiple insect species across diverse crop types, as it balances bias and variance and adapts to the natural variability in field- and laboratory-acquired images. The pipeline begins with image acquisition, where digital cameras collect input images under field and laboratory conditions. These raw images undergo pre-processing, including resizing, background removal, and cropping to isolate relevant insect-plant regions. The preprocessed images are then passed through an annotation step using tools like LabelImg, which generate XML-based bounding boxes around insect targets. Following annotation, the system performs feature extraction to capture visual descriptors, such as HOG for edge patterns, color histograms for color distribution, and shape descriptors for morphological characteristics. The extracted features are then subjected to feature selection, a critical step to reduce dimensionality and enhance model efficiency. Here, multiple techniques are applied–filter (example, CFS), wrapper (example, RFE), and embedded (e.g., Lasso, Random Forest importance)–to retain only the most informative attributes. These selected features feed into three core machine learning models: SVM, RF, and GBM. Each model independently predicts the presence of an insect, and the outputs are combined using a majority-vote ensemble to generate a final classification decision. This ensemble strategy improves robustness and generalization by leveraging the strengths of different learning algorithms. The diagram encapsulates a comprehensive modular system that accurately identifies insects in real-world agricultural conditions.

SHAP and Permutation Feature Importance (PFI) provide complementary forms of interpretability for the model’s decisions. SHAP assigns each feature a Shapley value that quantifies its positive or negative contribution to a specific prediction, thereby providing both local explanations (why an individual image is classified as CPB, aphid, ladybird, or healthy) and global trends across the dataset; in our case, SHAP highlights that high-impact cues include HOG gradients and biologically meaningful shape descriptors such as compactness and circularity. PFI, on the other hand, evaluates feature reliability at the model level by permuting one feature at a time and measuring the resulting drop in classification performance; features that cause the largest degradation are deemed essential for accurate inference. By intersecting the top-ranked features from SHAP (explainable contribution) and PFI (robust performance sensitivity), our hybrid-XAI pipeline retains only those descriptors that are simultaneously interpretable and consistently necessary for prediction. This dual-XAI strategy transforms the classifier from a black box into a transparent decision system, clarifying which visual/statistical traits drive insect recognition and ensuring stability under aggressive feature reduction.

Classification models and evaluation

Four classical classifiers, such as SVM, RF, KNN, and NB, are trained on the reduced feature sets. Evaluation is performed using stratified 5-fold cross-validation. Performance metrics include accuracy, precision, recall, F1-score, and computational latency to assess suitability for real-time edge deployment. Figure 4 illustrates a comprehensive pipeline that integrates XAI methods with traditional ML classifiers for insect classification. The process begins with the acquisition of raw image data, followed by feature extraction using pre-processing techniques.

Fig. 4

Overview of the proposed hybrid XAI-based feature selection and classification pipeline.

Two XAI techniques–SHAP and PFI–are applied in parallel to compute feature relevance scores. The top-k features from both methods are intersected to produce a reduced, high-quality feature set. This refined subset is then fed into a suite of classical classifiers (e.g., SVM, RF, KNN) for model training and evaluation. The outcome is a lightweight, interpretable detection system suitable for real-time deployment in agricultural settings that balances accuracy, efficiency, and explainability.

Rationale behind hybrid strategy

SHAP provides a local and global explanation by attributing feature importance based on Shapley values from cooperative game theory ¹². PFI, on the other hand, assesses feature relevance by quantifying how much the prediction error increases when a feature’s values are randomly permuted ⁵. The intersection of both ranks ensures that selected features are not only statistically significant but also interpretable and generalizable across classifiers.

Proposed hybrid CNN and feature selection pipeline logically structured

Figure 5 illustrates the logically structured workflow of the proposed hybrid CNN and feature-selection pipeline for insect classification. On the left, labeled input images (aphids, ladybird, CPB, and healthy potato leaf) are resized to $224\times 224\times 3$ and fed into a CNN feature extractor composed of stacked convolution$+$ReLU blocks and max-pooling layers, progressively transforming the image into deeper feature maps (e.g., $224\!\times \!224\!\times \!64 \rightarrow 112\!\times \!112\!\times \!128 \rightarrow 56\!\times \!56\!\times \!256 \rightarrow 28\!\times \!28\!\times \!512 \rightarrow 14\!\times \!14\!\times \!512 \rightarrow 7\!\times \!7\!\times \!512$). These high-level representations are flattened and passed through fully connected layers to produce a compact feature vector. In the middle, this feature vector is routed into a hybrid selection/optimization stage (shown as the “Feature Selection / Optimize Loss” module), where the most informative and non-redundant features are retained to enhance interpretability and efficiency. Finally, the selected features drive the softmax classifier to generate the output labels (Leptinotarsa decemlineata, Aphidoidea, Coccinellidae, or potato leaf), demonstrating an end-to-end design that couples deep representation learning with explainable feature pruning for robust and transparent insect recognition.

Fig. 5

Overview of the proposed CNN-driven hybrid feature selection and classification architecture.

Mathematical model of proposed hybrid systems

The proposed hybrid model sequentially integrates SVM, RF, and CNN, combining their strengths to enhance classification performance. Figure 6 shows the workflow of an insect detection and classification system applied to crops. It begins with input images from an agricultural dataset, including test images of insect-infested leaves. These images were first pre-processed using RGB color separation, shape analysis, diffusion, and filtering techniques to enhance image quality and segment the regions of interest related to the insect. This initial stage ensures that the insect features are highlighted, improving the system’s ability to detect and distinguish different insect species across varied backgrounds and environmental conditions. The feature extraction stage employs several machine learning techniques–HOG (Histogram of Oriented Gradients), Artificial Neural Networks (ANN), and Deep Neural Networks (DNN)–to derive robust feature vectors from preprocessed images. These features are then fed into a hybrid classification model that integrates SVM, RF, and CNN. The CNN component specifically handles the final classification and segmentation of the insects into categories: CPB, aphid, ladybird, and a generic class ’Feature X’ for unclassified or new features. The model outputs segmented regions and class labels, enabling precise pest identification for better management.

Fig. 6

Architecture of the proposed feature selection-driven hybrid CNN model for explainable detection of pests and beneficial insects in agricultural environments.

Feature extraction and vector formation are defined in Eq. (8). Given an input image I, preprocessing yields a feature set $\textbf{F} = \{f_1, f_2, \ldots , f_n\}$, where each $f_i$ represents extracted features like HOG descriptors, color histograms, and shape descriptors:

$$\begin{aligned} \textbf{F} = \Phi (I) \end{aligned}$$

(8)

where $\Phi (\cdot )$ denotes the feature extraction operation. SVM classifier is defined a linear or nonlinear decision function is formulated as in Eq. (9).

$$\begin{aligned} f(\textbf{x}) = \textbf{w}^\top \textbf{x} + b \end{aligned}$$

(9)

where $\textbf{w}$ and b are the weight vector and bias term, respectively, and $\textbf{x}$ is the input feature vector. The classification decision is defined as in Eq. (10).

$$\begin{aligned} \text {sign}(f(\textbf{x})) = {\left\{ \begin{array}{ll} +1, & \text {if } f(\textbf{x}) \ge 0 \\ -1, & \text {if } f(\textbf{x}) < 0 \end{array}\right. } \end{aligned}$$

(10)

Random forest classifieruses an ensemble of T decision trees as defined in Eq. (11).

$$\begin{aligned} h_t(\textbf{x}), \quad t=1,2,\ldots ,T \end{aligned}$$

(11)

The final prediction is given by majority vote as defined in Eq. (12).

$$\begin{aligned} H(\textbf{x}) = \text {mode}\left( \{h_1(\textbf{x}), h_2(\textbf{x}), \ldots , h_T(\textbf{x})\}\right) \end{aligned}$$

(12)

CNN-based segmentation and classification are defined in Eq. (13). For image-based classification, a CNN processes input feature maps $\textbf{X}$ through multiple convolutional layers:

$$\begin{aligned} \textbf{X}^{(l)} = \sigma \left( \textbf{W}^{(l)} * \textbf{X}^{(l-1)} + \textbf{b}^{(l)} \right) \end{aligned}$$

(13)

where $*$ denotes convolution, $\sigma (\cdot )$ is the activation function (e.g., ReLU), and $\textbf{W}^{(l)}$, $\textbf{b}^{(l)}$ are the weights and biases of layer l.

Hybrid model output via majority voting: The final classification output $\hat{y}$ is determined by combining the outputs of SVM, RF, and CNN as defined in Eq. (14).

$$\begin{aligned} \hat{y} = \text {mode}(\{ y_{\text {SVM}}, y_{\text {RF}}, y_{\text {CNN}} \}) \end{aligned}$$

(14)

where $\{ y_{\text {SVM}}, y_{\text {RF}}, y_{\text {CNN}} \}$ are the individual predictions. This mathematical formulation ensures robust performance by leveraging the complementary learning capabilities of SVM, RF, and CNN, tailored for insect detection and classification across diverse crop environments.

Evaluation metrics

Several standard evaluation metrics assess the performance of the proposed hybrid insect classification models, including accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (ROC-AUC). These metrics comprehensively understand the model’s correctness and robustness across different conditions and insect types. All metrics are calculated using a 10-fold cross-validation strategy to ensure reliability and reduce bias. Accuracy measures the model’s overall correctness and is defined in Eq. (15).

$$\begin{aligned} \text {Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} \end{aligned}$$

(15)

where TP, TN, FP, and FN are the true positives, true negatives, false positives, and false negatives, respectively. Precision evaluates how many positive predicted instances are defined in Eq. (16).

$$\begin{aligned} \text {Precision} = \frac{TP}{TP + FP} \end{aligned}$$

(16)

This is crucial in pest detection to avoid false alarms that may lead to unnecessary pesticide application. Recall quantifies the model’s ability to detect actual positives as defined in Eq. (17).

$$\begin{aligned} \text {Recall} = \frac{TP}{TP + FN} \end{aligned}$$

(17)

High recall ensures that most infested cases are identified, which is essential for timely pest management. The F1-score is the harmonic mean of precision and recall, offering a balanced measure when the dataset is imbalanced as defined in Eq. (18).

$$\begin{aligned} \text {F1-Score} = 2 \times \frac{\text {Precision} \times \text {Recall}}{\text {Precision} + \text {Recall}} \end{aligned}$$

(18)

The ROC-AUC score represents the model’s ability to distinguish between classes across all threshold levels. It is computed as the area under the ROC curve, which plots the true positive rate (Recall) against the false positive rate (FPR) as defined in Eq. (19).

$$\begin{aligned} \text {AUC} = \int _{0}^{1} \text {TPR}(FPR) \, dFPR \end{aligned}$$

(19)

We apply 10-fold cross-validation to ensure generalizability, dividing the dataset into 10 equal parts. The model is trained on 9 parts and tested on the remaining part. This process is repeated 10 times, each with a different test fold, and the average of all metric scores is reported to minimize overfitting and variance. These metrics provide a robust framework for evaluating the hybrid classification model’s predictive power and practicality across varied environmental and crop-specific conditions.

Criteria for selecting the top 11 features

Figure 7 shows a two-stage XAI-guided feature selection process: SHAP and PFI first rank features independently, and their top-ranked features overlap to form a robust hybrid candidate set. Accuracy is then evaluated as features are added, and the curve plateaus at 11 features, so k=11 is chosen as the smallest stable subset achieving near-maximum performance.

Fig. 7

Criteria for selecting the top 11 features using the proposed two-stage explainability-guided strategy. In Stage 1, SHAP and Permutation Feature Importance (PFI) are computed independently to produce ranked importance lists; the intersection of their top-ranked features forms a hybrid candidate set that retains only features consistently identified as influential by both XAI measures, thereby reducing redundancy and method-specific bias. In Stage 2, classification accuracy is evaluated over progressively larger hybrid subsets ($k=6$ to 20), revealing a rapid performance increase followed by a stable plateau; the optimal point occurs at $k=11$, where peak (or near-peak) accuracy is achieved with minimal feature count and high stability. Thus, the top 11 features are selected as the smallest compact subset that preserves maximum predictive performance while improving interpretability and computational efficiency.

To determine the optimal feature subset, we adopted a two-stage explainability-guided selection criterion. First, SHAP and Permutation Feature Importance (PFI) were computed independently on the trained base classifier to produce two ranked importance lists, where SHAP reflects each feature’s marginal contribution to predictions, and PFI quantifies the sensitivity of model performance to feature perturbation. We then formed a hybrid candidate set by taking the intersection of the top-k features from both rankings, retaining only those consistently identified as important by both XAI measures and thereby reducing redundancy and method-specific bias. Second, to select the final number of features, we evaluated classification accuracy under progressively larger hybrid subsets ($k=6$ to 20) using cross-validation. Accuracy increased sharply as features were added, reached its maximum (or near-maximum) at $k=11$, and then exhibited a stable plateau with negligible gains beyond this point, while larger subsets occasionally introduced minor fluctuations. Consequently, the top 11 features were selected as the smallest explainable subset achieving peak and stable performance, providing the best trade-off between interpretability, compactness, and accuracy.

Cross-validation protocol

Table 1 summarizes the cross-validation workflow adopted to ensure that the reported accuracy stability across feature subsets is statistically valid. A stratified 10-fold split was first generated, ensuring each fold preserved the original class balance and preventing bias toward any insect category. Importantly, the same fold partitions were reused across all subset sizes (e.g., $k=6$–20 and the top-11 set), ensuring that performance differences arise from the selected features rather than from random resampling. Within each fold, SHAP, PFI, and their hybrid intersection were computed strictly on the training split, and only the resulting features were applied to the held-out test split, eliminating information leakage. Accuracy was then aggregated as mean ± standard deviation over folds to capture both central performance and variability. Finally, stability was confirmed when accuracy gains beyond the top-11 subset remained smaller than the cross-fold variance and within one standard deviation of the best larger subsets, supporting a statistically consistent plateau and validating robustness of the compact feature set.

Table 1 Cross-validation protocol used to ensure statistically valid accuracy stability across feature subsets.

Experimental setup

Experiments were conducted on a curated dataset comprising images of both harmful (e.g., Colorado Potato Beetle, aphids) and beneficial (e.g., ladybirds) insect species. We extracted a diverse set of handcrafted and CNN-derived features and evaluated four lightweight classifiers (SVM, RF, KNN, NB) under various feature selection strategies. To validate the effectiveness of the proposed hybrid model for insect detection in multiple crop types, a series of experiments were conducted using the curated dataset. The dataset was split into training and test sets using a 10-fold cross-validation strategy to ensure robustness and reduce variance. Feature selection techniques were applied before model training to identify the most informative features, and three machine learning classifiers, such as SVM, RF, and GBM were trained using these features. Their predictions were aggregated using a majority-vote ensemble to produce the final classification. Experiments were performed under various lighting and environmental conditions to evaluate model generalization in field images. Performance was assessed using standard evaluation metrics, including accuracy, precision, recall, F1-score, and ROC-AUC, providing a comprehensive measure of the model’s effectiveness in real-world agricultural monitoring scenarios.

A brief computational complexity and model-parameter analysis was included to emphasize the efficiency of the proposed pipeline. Let p denote the original feature dimensionality and k the reduced dimensionality after hybrid SHAP+PFI selection. Since SHAP and PFI are computed offline during training, their cost does not affect real-time deployment; the main benefit appears at inference, where the prediction cost drops from approximately O(p) to O(k) with $k \ll p$. In our experiments, the hybrid XAI method reduced the feature space by about $63\%$, selecting $k=11$ features while preserving near-maximum accuracy. This reduction lowered both model size and latency across all classifiers: for example, Random Forest required roughly 1000 effective parameters and achieved 0.6 s prediction time, while SVM used about 850 parameters with 0.5 s prediction time under the proposed selection, outperforming conventional selectors that retained larger k and incurred higher latency. Overall, the complexity analysis confirms that the proposed explainability-guided selection yields a compact, computationally efficient model suitable for edge deployment without sacrificing predictive performance.

Transfer learning was used in this study, leveraging a pre-trained CNN backbone to enhance feature extraction. Specifically, we employed MobileNetV2, pre-trained on ImageNet, as a fixed feature extractor: the original classification head was removed, and deep embeddings were extracted from the final global average pooling layer. These CNN-derived features were concatenated with handcrafted descriptors (HOG, color histograms/statistics, and shape features) to form the complete feature pool. The proposed SHAP+PFI hybrid feature selection was then applied to this combined set to retain only the most informative and non-redundant features before training the lightweight classifiers and hybrid ensemble. To reduce overfitting and preserve edge-deployment feasibility, the pre-trained CNN weights were not fully fine-tuned; instead, transfer learning was used only in feature-extraction mode.

Results

Experimental evaluation

Figure 8 presents a comparative performance analysis of four classifiers across standard metrics. The Hybrid model consistently achieves the best results, attaining an accuracy of $96.7\%$, precision of $96.5\%$, recall of $96.9\%$, and an F1-score of $96.7\%$, confirming its strong and balanced detection capability. Among individual models, GBM performs second best with $95.0\%$ accuracy, $94.7\%$ precision, $95.1\%$ recall, and $94.9\%$ F1-score, indicating robust generalization but still trailing the ensemble.SVM yields competitive performance (accuracy $94.1\%$, precision $93.8\%$, recall $94.2\%$, F1-score $94.0\%$), slightly outperforming RF, which records $93.2\%$ accuracy, $92.5\%$ precision, $93.0\%$ recall, and $92.7\%$ F1-score. Overall, the figure demonstrates that combining complementary learners in the Hybrid approach provides a clear gain of roughly 1.7–3.5 percentage points over single classifiers, validating the effectiveness of the proposed hybrid pipeline for accurate and reliable insect classification.

Fig. 8

Experimental evaluation of the proposed hybrid framework against baseline classifiers. Grouped bars report Accuracy, Precision, Recall, and F1-score (%) for SVM, RF, GBM, and the Hybrid model.

Figure 9 shows that the proposed hybrid model achieves strong and well-balanced class-wise performance. CPB samples are correctly classified at $96.8\%$, with small confusion into aphid ($1.6\%$), ladybird ($0.8\%$), and healthy ($0.8\%$). Aphids obtain a true-positive rate of $96.0\%$, with minor misclassification mainly into CPB ($2.0\%$) and ladybird ($1.2\%$). Ladybirds are correctly recognized at $96.4\%$, with limited confusion into CPB ($1.2\%$) and aphid ($1.6\%$). Healthy leaves show the highest separability, reaching $97.2\%$ correct classification, and only $\le 1.2\%$ confusion into insect classes. Overall, the diagonally dominant matrix confirms that the high overall accuracy is supported by consistently high recall across both pest and beneficial insect categories, with no class suffering from systematic bias.

Fig. 9

Normalized confusion matrix (%) of the proposed hybrid model for four classes: Colorado potato beetle (CPB), aphid, ladybird, and healthy leaf. Rows denote true labels and columns denote predicted labels.

Figure 10 visualizes class-wise separability using one-vs-rest ROC analysis. All four ROC curves lie near the upper-left corner, indicating high true-positive rates at very low false-positive rates. Consistent with the strong overall performance of the model, the areas under the curves (AUCs) are close to unity in this demonstration, with each class achieving an AUC of approximately 0.99 (CPB), 0.99 (aphid), 0.99 (ladybird), and 0.99 (healthy leaf). The pronounced gap above the chance line confirms that the classifier maintains excellent discrimination across pest, beneficial, and healthy categories, supporting the reported high accuracy and balanced class-wise recognition.

Fig. 10

One-vs-rest multi-class ROC curves of the proposed hybrid insect classifier for four classes (CPB, aphid, ladybird, and healthy leaf). Each curve is plotted by treating one class as positive and the remaining classes as negative; the dashed diagonal denotes chance-level performance.

Table 2 reports the stratified 10-fold cross-validation accuracy (mean ± SD) for different feature selection methods as the number of retained features increases. The proposed Hybrid XAI (SHAP+PFI) method shows the fastest improvement and the highest stability: accuracy rises from $65.0\%\pm 2.1$ at $k=6$ to $82.0\%\pm 1.8$ at $k=9$, reaching a clear plateau of $\mathbf {90.0\%\pm 1.2}$ by $k=11$ and maintaining essentially the same performance up to $k=20$ with decreasing variance. In contrast, conventional methods (PCA, MI, Chi2, Fisher, MIC, and VarThresh) achieve substantially lower peak accuracies ($\sim$60–72%) and exhibit either weak gains or noticeable drops as more features are added (e.g., PCA and MI decline beyond $k=11$). The relatively larger SDs for baselines indicate higher fold-to-fold fluctuation, whereas the smaller SDs of the Hybrid XAI curve confirm that its plateau at the top-11 subset is statistically consistent, supporting the claim that a compact, explainable feature set preserves high accuracy without instability from redundant features.

Table 2 Stratified 10-fold CV accuracy (mean ± SD, %) across feature subset sizes.

Identification region of interest and insect segmentation

In this study, the primary objective of insect segmentation is to accurately separate insects from the surrounding background in captured images. Figure 11 illustrates the process of insect segmentation and region of interest (ROI) identification in different crop images, focusing on three key insect types: the Colorado potato beetle, the seven-spot ladybird, and aphids.

Fig. 11

The process of insect segmentation and region of interest (ROI) identification in different crop images, focusing on three key insect types: the Colorado potato beetle, the seven-spot ladybird, and aphids.

Initially, images are converted into multiple color spaces, such as RGB and shape-based representations, to emphasize various visual characteristics of the insects. A diffusion-based filtering technique enhances region boundaries and ensures that insect features are more distinguishable. Finally, a precise region of interest (ROI) is defined around each segmented insect, effectively isolating it from the background and capturing all its body parts. This ROI extraction facilitates accurate classification in the subsequent stages of the model and provides critical information for downstream analysis. The first row shows the original input images captured in field conditions, presenting the natural appearance of crop plants infested with insects. The second row displays the segmented regions, where insect features have been separated from the complex background to improve the clarity of subsequent analysis. The identified ROIs are marked in the third row, showing how the insect bodies have been isolated for further classification. This process involves applying various image processing techniques, including color-space separation, shape-based filtering, and morphological operations, to differentiate insect structures from surrounding foliage and soil textures. The final row of images for each insect type provides detailed examples of the segmentation and identification outcomes. The segmentation for the Colorado potato beetle and the seven-spot ladybird captures fine morphological details such as stripes and spots. At the same time, aphids are detected as smaller, distinct white regions in the ROI masks. By isolating these insect regions, the approach ensures precise input for classification models, which is crucial for accurate detection of pests and beneficial insects. This segmentation and ROI identification process is critical in reducing background noise and highlighting the key insect features essential for robust machine learning-based pest management systems.

Figure 12 illustrates the pixel intensity distribution for two classes in an image: the background (solid blue line) and the foreground (insect) (red dashed line). The x-axis represents the pixel intensity values ranging from 0 (black) to 255 (white), while the y-axis indicates the frequency of those intensity values across the image. The background dominates the picture, as noted in the high-frequency peak centered at an intensity of 190–200 pixels, with a maximum frequency exceeding 4000 pixels. This indicates that the background is mainly composed of brighter pixels, which is common in images where the subject (insect) is placed on a well-lit or white surface.

Fig. 12

The pixel intensity distribution for two classes in an image: background (blue solid line) and foreground (insect) (red dashed line).

In contrast, the foreground (insect) has a much lower frequency distribution across all pixel intensities. The red dashed line is relatively flat and remains below 50 pixels at all intensities, confirming that the insect occupies a small portion of the image, likely around 5% or less. The insect’s pixel intensity is more uniformly distributed, with slight variation in the 40–120 range, which typically corresponds to darker tones. This pixel count and intensity disparity highlight the segmentation challenge – extracting a small, low-contrast object from a dominant, high-intensity background.

Feature selection impact/XAI analysis

Application of feature selection reduced dimensionality by 40-60% without sacrificing model performance. Embedded methods offered the best balance of computational cost and classification accuracy. Figure 13 illustrates the correlation matrices for two distinct feature sets used in classifying insect classes: shape features (left panel) and texture features (right panel). Each matrix is a visual representation of the pairwise relationships between the extracted features, with color gradients indicating the strength and direction of these correlations. Positive correlations, represented by warm colors (yellow to red), suggest that two features tend to increase or decrease together. In contrast, cool colors (blue shades) denote negative correlations, where an increase in one feature corresponds to a decrease in the other. The diagonal line of each matrix, marked in bright color, represents the perfect correlation of each feature with itself (value = 1). By analyzing these matrices, we can observe that shape features such as circularity, roundness, and solidity exhibit strong positive correlations, implying they often co-occur in certain insect classes. Conversely, aspect ratio and elongation negatively correlate with circularity and solidity, highlighting their distinct contributions to shape characterization. Similarly, features like contrast, correlation, and energy in the texture matrix show notable relationships that can be leveraged to enhance insect classification. This comprehensive view helps identify feature redundancies and complementarities, ultimately informing the feature selection strategy to boost model performance in insect detection tasks.

Fig. 13

The correlation matrices for two distinct feature sets used in classifying insect classes: shape features (left panel) and texture features (right panel).

Table  3 thoroughly evaluates 29 different features. Each entry lists the feature’s rank, name, and its mean weight alongside the standard deviation (SD), expressed in the format mean ± SD. The most influential feature is hog_1757 with a mean contribution of $0.030 \pm 0.020$, followed by Compactness at $0.025 \pm 0.018$ and hog_611 at $0.021 \pm 0.022$. These values indicate that these features significantly impact the model’s performance, albeit with some variability. Several shape and geometric descriptors also rank highly in the list. Circularity contributes $0.022 \pm 0.017$ to the model, while aspect ratio and roundness contribute $0.019 \pm 0.014$ and $0.018 \pm 0.013$, respectively. The feature labeled Alr (possibly representing asymmetry or angular length ratio) also holds a solid rank with a contribution of $0.019 \pm 0.008$, indicating a strong yet consistent effect on the model’s predictions. This demonstrates the significance of incorporating texture and shape descriptors for improved model interpretability. In the mid-ranking positions, features such as hog_1643, hog_1266, hog_60, and hog_298 show weights ranging from 0.015 to 0.016, highlighting the relevance of HOG features across various indices.

Table 3 Ranking of features based on their contribution to the model’s performance. Values are expressed as mean ± standard deviation.

Complementary features like Std_hist ($0.016 \pm 0.017$), Hollowness ($0.015 \pm 0.012$), and Texture Entropy ($0.015 \pm 0.011$) demonstrate that both statistical and structural descriptors contribute meaningfully, even when not the top-ranked. Towards the bottom of the list, features like Texture Energy ($0.012 \pm 0.009$), Texture Correlation ($0.011 \pm 0.008$), and Texture Variance ($0.010 \pm 0.007$) offer lower but stable contributions. This indicates that while they provide useful information, their influence is limited compared to higher-ranked features. Meanwhile, additional HOG features such as hog_504, hog_598, and hog_206 show moderate weights of around 0.013, suggesting that not all gradient-based features are equally valuable. The table highlights the benefit of using a diverse set of features to enhance the predictive performance and robustness of the model.

Figure 14 presents a comprehensive analysis of feature importance and individual feature contributions for a random forest model in insect classification tasks. It consists of two subplots. Figure 14a is a horizontal bar chart displaying the average feature importance scores of the top ten features (measured by Shapley values). Figure 14b is a SHAP summary plot illustrating the influence of these features on individual model predictions. These visualizations help understand which features drive the model’s decisions most strongly and how they vary across samples. Figure 14a bar chart highlights that the arealength ratio (alr) feature has the highest average importance, with a SHAP value of 0.030. It is followed by roundness (0.025) and length perimeter ratio (lpr) at 0.023. Features like compactness (0.022), circularity (0.020), and standard deviation histogram (std_hist, 0.018) also contribute significantly to the model’s performance. At the lower end, features such as energy (0.015), rectangularity (0.014), hollowness (0.013), and hog_889 (0.012) exhibit minimal influence, suggesting that their impact on classification decisions is less critical.

Fig. 14

A comprehensive analysis of feature importance and individual feature contributions for a random forest model in insect classification tasks.

The SHAP summary Fig. 14b delves deeper by showing how individual feature contributions vary across the dataset. Each feature’s SHAP values are represented by colored dots, with colors indicating the magnitude of feature values (blue for low values and red for high values). For instance, alr and roundness features generally have positive SHAP values, suggesting that higher feature values push the model’s predictions upwards. In contrast, features like hollowness and hog_889 have SHAP values closer to zero, confirming their lower influence on individual predictions. The clear separation in feature impacts revealed by these plots underscores the importance of robust feature selection in enhancing the random forest model’s performance. Features with higher average SHAP values, like alr and roundness, can be prioritized for training to improve accuracy. In contrast, less impactful features like hollowness and hog_889 may be omitted or further evaluated. The visual insights also highlight the model’s interpretability, revealing which morphological and textural features are most informative for distinguishing between insect classes. Future work may include validating these insights across different insect datasets or exploring how these feature-importance patterns generalize to other agricultural settings.

Figure 15 compares global feature relevance estimated by SHAP and PFI, highlighting their agreement and the resulting hybrid selection. The most influential feature is hog_1757, scoring 0.030 under SHAP and 0.028 under PFI, followed by Compactness (0.025 SHAP, 0.024 PFI) and hog_611 (0.021 SHAP, 0.020 PFI). Key shape descriptors such as Circularity (0.022 vs. 0.021), Aspect Ratio (0.019 vs. 0.018), ALR (0.019 vs. 0.020), Roundness (0.018 vs. 0.017), and Elongation (0.017 vs. 0.016) show consistently high contributions across both XAI tools, confirming their biological relevance. Statistical and texture-related features such as Std_hist (0.016 vs. 0.015), hog_1643 (0.016 vs. 0.014), and Texture Entropy (0.015 vs. 0.016) provide complementary discrimination. The dashed boxes mark the top-11 overlapping features, demonstrating strong SHAP–PFI concordance and justifying the compact hybrid subset that achieves near-peak accuracy while remaining interpretable.

Fig. 15

Dual-XAI feature importance comparison using SHAP and Permutation Feature Importance (PFI). Bars show normalized importance scores for the top-ranked features, while dashed boxes indicate the hybrid overlap (intersection) retained as the final compact subset (top-11).

Figure 16 provides a qualitative and quantitative view of how the most informative features drive the classifier’s decisions. The top-ranked features exhibit the highest mean absolute SHAP contributions, confirming their dominant role in separability. For example, the strongest influence is observed for hog_1757 (mean SHAP $\approx 0.030$), followed by Compactness ($\approx 0.025$), hog_611 ($\approx 0.021$), Circularity ($\approx 0.022$), and hog_1294 ($\approx 0.020$). Morphological descriptors such as Aspect Ratio ($\approx 0.019$), ALR ($\approx 0.019$), Roundness ($\approx 0.018$), and Elongation ($\approx 0.017$ also show consistently high positive/negative Shapley spreads, indicating that variation in insect body shape and symmetry strongly shifts model confidence. In contrast, lower-ranked texture descriptors (e.g., Texture Energy and Texture Correlation, SHAP $\le 0.012$) cluster near zero, demonstrating minimal effect on predictions. Overall, the SHAP distribution confirms that the hybrid-XAI pipeline prioritizes biologically meaningful gradient and shape cues, and explains why the model retains near-peak accuracy using the compact top-11 feature subset.

Fig. 16

SHAP summary visualization for the proposed hybrid classifier (one-vs-rest), showing the distribution of Shapley values for each feature with respect to the target class. Features are ordered by global importance (mean absolute SHAP), where larger absolute values indicate stronger influence on the prediction.

Table 4 shows that the proposed model focuses primarily on a small, biologically meaningful set of visual-gradient and morphological descriptors. The strongest contributors are HOG-based features (hog_1757, hog_611, hog_1294; weights up to $0.030\pm 0.020$), indicating that edge orientation and stripe/spot boundaries are critical for distinguishing species under variable field backgrounds. Immediately after, shape metrics such as Compactness ($0.025\pm 0.018$), Circularity ($0.022\pm 0.017$), Aspect Ratio ($0.019\pm 0.014$), ALR ($0.019\pm 0.008$), Roundness ($0.018\pm 0.013$), and Elongation ($0.017\pm 0.012$) dominate the ranking, confirming that insect body geometry and symmetry provide stable class-discriminative cues (e.g., round ladybird adults versus more elongated CPB or aphid forms). Finally, a small number of statistical features–Std_hist ($0.016\pm 0.017$) and Texture Entropy ($0.015\pm 0.011$)–capture color/brightness spread and surface granularity, helping separate visually distinct taxa (red-spotted ladybirds vs. pale aphid colonies) and insect regions from healthy leaves. The convergence of SHAP and PFI on this subset explains why high accuracy is retained with only 11 features, as these cues are both visually robust and entomologically interpretable.

Table 4 Most influential visual and statistical features used by the proposed model for insect classification, as jointly identified by SHAP and PFI (top-ranked subset).

Model performance

Figure 17 presents a comparative analysis of five machine learning models–SVM, Random Forest, GBM, CNN, and the Hybrid Model–evaluated across four performance metrics: Accuracy, Precision, Recall, and F1-Score. Each model is represented on the x-axis, with grouped bars illustrating metric values. The Hybrid Model shows the highest overall performance, reaching an average accuracy of 0.95, precision of 0.94, recall of 0.93, and F1-Score of 0.94. These results indicate that the hybrid model outperforms the other models in all four metrics. The CNN model also performs strongly, with an accuracy of 0.93, a precision of 0.92,a recall of 0.91, and an F1-score of 0.92, showcasing the potential of deep learning for robust insect detection. In contrast, SVM shows the lowest performance among the models, with an accuracy of 0.88, a precision of 0.85, a recall of 0.87 and an F1-Score of 0.86. Meanwhile, the Random Forest and GBM models demonstrate competitive performance, achieving metric values of around 0.89–0.91, suggesting they provide a good balance of performance for practical applications. Figure 17 also shows clear grouping and color-coding, revealing that while traditional models like SVM and Random Forest offer reliable results, advanced techniques like CNN and Hybrid models significantly boost accuracy and generalization. This Fig. 17 illustrates how integrating advanced feature representations and ensemble learning can improve insect classification performance, which is crucial for precision agriculture and integrated pest management. These insights reinforce the importance of using advanced machine learning strategies, particularly CNN and hybrid model approaches, for improved pest and beneficial insect classification.

Fig. 17

A comparative analysis of five machine learning models evaluated across four performance metrics.

The hybrid model achieved an overall accuracy of 96.7%, with SVM and GBM contributing most significantly. Table 5 summarizes the comparative results. The hybrid model demonstrates the highest overall performance, achieving 96.7% accuracy, 96.5% precision, 96.9% recall, and an F1-score of 96.7%. In contrast, the SVM model attains 94.1% accuracy and 94.0% F1-score, slightly outperforming RF, which achieves 93.2% accuracy and 92.7% F1-score. The GBM model also performs well, with an accuracy of 95.0% and an F1-score of 94.9%, indicating its robustness in classification tasks. These results show that while traditional machine learning models (SVM, RF, and GBM) exhibit strong, consistent performance, the proposed hybrid model significantly outperforms them across all evaluated metrics, accurately identifying relevant instances, with a precision of 96.5%, indicating reliable predictions. Specifically, the hybrid model’s recall of 96.9% suggests its superior ability to accurately identify relevant instances, while its precision of 96.5% indicates reliable predictions with minimal false positives. This comprehensive performance improvement underscores the effectiveness of the hybrid approach in integrating multiple classifiers’ strengths to achieve higher accuracy, reliability, and robustness in insect classification.

Table 5 Model Performance Comparison.

Figure 18 provides a comprehensive comparative performance analysis of four machine learning classifiers–Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Naïve Bayes (NB)–across seven feature selection techniques: PCA, Mutual Information (MI), Chi-squared (Chi2), Fisher separation criterion (Fisher), Variance Thresholding (VarThresh), Maximal Information Coefficient (MIC), and the Proposed Method. Each subplot contains bar plots illustrating four critical metrics: precision, recall, F1-score, and accuracy. This side-by-side comparison of classifiers and techniques reveals how feature selection methods impact classifier performance. The Random Forest classifier consistently achieved the highest scores across most feature selection techniques. For the Proposed Method, RF reached peak values: precision of 0.90, recall of 0.91, F1-score of 0.89, and accuracy of 0.92–highlighting the synergy between this classifier and the newly developed feature selection method. Similarly, SVM also performed exceptionally well with the Proposed Method, achieving accuracy of 0.90. In contrast, traditional feature selection techniques, such as VarThresh and Fisher, yielded lower performance for all classifiers, with average accuracy values of around 0.82 for RF and 0.80 for SVM. Among the classifiers, Naïve Bayes consistently exhibited lower metrics across all feature selection methods, with accuracy peaking at 0.85 (Proposed Method) but dropping as low as 0.75 (VarThresh). The KNN classifier demonstrated moderate improvements with the Proposed Method, achieving an accuracy of 0.87 compared to 0.77 with VarThresh. These findings emphasize that while the Proposed Method significantly boosts performance for all classifiers, the RF classifier benefits the most, likely due to its ensemble nature and robustness to high-dimensional feature sets. This comprehensive figure thus highlights the relative strengths of classifiers and the critical role of tailored feature selection strategies in improv.

Fig. 18

A comparative analysis of five machine learning models evaluated across four performance metrics.

Figure 19 illustrates the impact of progressively adding features (from 6 to 20) on the accuracy of insect classification models across seven feature selection techniques.

Fig. 19

The impact of progressively adding features (from 6 to 20) on the accuracy of insect classification models across seven feature selection techniques.

Notably, the proposed XAI method (red line) demonstrates superior stability and performance compared to other methods. As seen in the figure, the accuracy for the XAI method rapidly increases from 65% at 6 features to a peak accuracy of 90% by 11 features. After reaching this point, the XAI method maintains an accuracy plateau at 90%, showing remarkable stability even as more features are added. In contrast, conventional methods like PCA and MI initially show improvement, with PCA achieving approximately 72% accuracy at 11 features and MI achieving the same accuracy. However, these traditional methods exhibit fluctuations beyond 11 features: PCA drops to 66% at 15 features, while MI decreases to 61% at 20 features, indicating a lack of robustness when more features are included. This highlights the tendency of these methods to become unstable or susceptible to noise when additional features are considered. Figure 19 also highlights the role of individual feature importance: HOG features contribute significantly to early improvements, with the first three HOG features increasing the XAI method’s accuracy to 65%.

Table 6 provides a comprehensive comparison of the classification accuracy of four machine learning models–Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbour (KNN), and Naïve Bayes (NB)–across several feature selection techniques.

Table 6 A comparative analysis of the accuracy and computational complexity of the proposed approach versus conventional feature selection methods for insect classification is presented.

For instance, the All+PCA feature selection method achieved accuracies of 85% (SVM), 87% (RF), 83% (KNN), and 80% (NB). In the Mutual Information (MI) method, accuracies improved slightly, with SVM and RF reaching 86% and 88% respectively, while KNN and NB achieved 84% and 81%. These results indicate a consistent advantage of RF over other classifiers across different feature selection methods. Regarding computational complexity, the number of model parameters ranged from 500 to 1,500, depending on the feature selection method and classifier. For example, the RF classifier consistently had the highest parameter count across techniques (e.g., 1500 in All+PCA and 1400 in MI), while NB had the fewest (e.g., 500 in All+PCA and 450 in MI). This highlights the trade-off between model complexity and performance, as RF generally outperformed other classifiers at the expense of increased parameter count. Training and prediction times also show apparent differences. For All+PCA, RF required 15 seconds to train and 1.0 seconds to predict, whereas NB only needed 8/0.5 seconds. These time differences suggest that RF offers better accuracy but requires more computational resources. The Variance Thresholding method exhibited the fastest overall training/prediction times for NB (5/0.3 seconds) and KNN (7/0.4 seconds), albeit with slightly lower accuracy than other methods. The proposed XAI method achieved a standout performance, which balanced high accuracy with low computational demands. Here, RF achieved an impressive 92% accuracy, outperforming all other techniques, with only 1000 parameters and 9/0.6 seconds of training and prediction time. Similarly, SVM and KNN achieved 90% and 87% accuracy, respectively, while maintaining relatively low model complexity and time requirements. The results underscore the XAI method’s potential to deliver robust insect classification performance without excessive computational overhead, making it a practical and scalable approach for real-world agricultural monitoring applications. Subsequent inclusion of colour features (entropy_hist, energy_hist, std_hist) and shape features (elongation, ferrat major axis) propels accuracy to 90%. In contrast, methods like the Fisher separation criterion and Chi-squared coefficient hover around 60-65% after 20 features, reflecting their weaker performance compared to the proposed XAI approach. This Fig. 19 underscores the effectiveness of the proposed XAI method in consistently selecting the most informative features while minimizing noise-related fluctuations, achieving both stability and high performance.

Comparison with Baseline methods

The comparative analysis with baseline methods is essential to validate the effectiveness and practical value of the proposed hybrid XAI-based feature selection strategy. Traditional feature selection techniques–such as filter (Chi-square, ReliefF), wrapper (Recursive Feature Elimination), and embedded (LASSO) approaches–are commonly used but often focus solely on statistical significance or model performance, without providing interpretability. Moreover, using all features without selection can introduce noise and redundancy, reducing classification accuracy and increasing computational cost. By benchmarking our hybrid approach against established methods, we demonstrate that it offers not only superior performance across key metrics (accuracy, precision, recall, and F1-score) but also enhances model transparency and feature relevance. This evaluation confirms the added value of explainable AI in developing interpretable, efficient insect classification systems for smart agricultural applications. Figure 20 presents a comparative analysis of four standard performance metrics–Accuracy, Precision, Recall, and F1-score–across five different feature selection techniques: All Features (no selection), Chi-Square (Chi2), Recursive Feature Elimination (RFE), Lasso (embedded), and the proposed Hybrid XAI-based method combining SHAP and PFI.

Fig. 20

Comparison of performance metrics (Accuracy, Precision, Recall, F1-score) across different feature selection methods: No Feature Selection, Chi-square, ReliefF, Recursive Feature Elimination (RFE), LASSO, and the proposed Hybrid XAI-based approach (SHAP + PFI). The Hybrid method consistently outperforms the baselines, demonstrating enhanced classification performance and robustness.

Each group of bars represents one method, with individual bars within the group denoting the value of each metric. For the “All Features” approach, the performance was the lowest across all metrics (Accuracy: 81.2%, Precision: 80.5%, Recall: 79.8%, F1-score: 80.1%), highlighting the limitations of using unfiltered high-dimensional data. In contrast, the proposed Hybrid (SHAP + PFI) method demonstrated the highest performance, with an Accuracy of 89.1%, Precision of 88.7%, Recall of 88.9%, and F1-score of 88.8%, significantly outperforming all baselines. Even compared to strong embedded methods like Lasso (Accuracy: 86.5%, F1-score: 85.8%) or wrapper methods like RFE (Accuracy: 85.3%, F1-score: 84.9%), the hybrid approach provides an average performance gain of  3–4% across all metrics. This improvement confirms that integrating SHAP and PFI not only enhances the predictive power of classical classifiers but also ensures the selected features are both informative and interpretable–ideal for real-time, field-deployable insect classification systems.

Figure 21 illustrates a classifier-wise comparison of classification accuracy across multiple feature selection methods, including Chi-square (Chi2), ReliefF, Recursive Feature Elimination (RFE), LASSO, and the proposed Hybrid XAI-based approach (SHAP + PFI). For the Support Vector Machine (SVM) classifier, the hybrid method achieves the highest accuracy at 94.2%, compared to 90.5% with RFE, 89.1% with Chi2, and 88.3% with LASSO. Similarly, for the Random Forest (RF) classifier, the hybrid method again leads with 96.7%, outperforming ReliefF (93.8%), Chi2 (92.1%), and LASSO (91.6%). This highlights the hybrid approach’s ability to select highly informative features that enhance discriminative power. For the K-Nearest Neighbors (KNN) classifier, which is more sensitive to feature quality, the hybrid method achieves an accuracy of 91.3%, compared to 87.4% with RFE and 86.0% with Chi2. Across all classifiers, the consistently superior performance of the hybrid feature selection technique demonstrates its robustness and adaptability. These results validate the importance of integrating model interpretability (via SHAP) and statistical robustness (via PFI) to improve both predictive accuracy and feature relevance.

Fig. 21

Classifier-wise comparison of Accuracy across various feature selection methods, including Chi2, ReliefF, RFE, Lasso, and the proposed Hybrid (SHAP+PFI). The hybrid method consistently outperforms baselines across SVM, RF, and KNN classifiers, underscoring its effectiveness in enhancing classification performance.

Performance metrics

We evaluated each method using accuracy, F1-score, and model latency on a Raspberry Pi 4 Model B (as a proxy for low-cost edge deployment). Our hybrid method consistently outperformed others in accuracy while reducing the number of features by over 60%. The table  7 presents a comprehensive evaluation of various feature selection strategies based on four metrics: Accuracy, F1-score, Latency, and Feature Reduction percentage.

Table 7 Performance Comparison of Feature Selection Methods on Raspberry Pi 4 (Edge Deployment).

Among the methods assessed–No Feature Selection, Chi2, ReliefF, Recursive Feature Elimination (RFE), LASSO, and the proposed Hybrid SHAP+PFI approach–the hybrid method achieved the highest classification accuracy of 94.2% and the best F1-score of 0.93, outperforming all baselines. In contrast, using all features without selection yielded lower accuracy (88.4%) and an F1-score of 0.87, underscoring the inefficiency of high-dimensional feature spaces with redundant or irrelevant attributes. In addition to improved predictive performance, the hybrid method demonstrated practical advantages for edge deployment. It achieved the lowest latency (1.62 seconds) for inference on a Raspberry Pi 4 Model B, compared to 2.35 seconds with no feature selection and 2.20 seconds with RFE. Importantly, the hybrid approach also delivered the highest feature reduction rate of 63%, substantially decreasing the model complexity without sacrificing performance. These results validate the hybrid XAI-based strategy as both an effective and efficient solution, particularly suited for real-time insect classification in resource-constrained environments such as smart farms. The proposed SHAP+PFI hybrid feature selection strategy demonstrated clear advantages across multiple dimensions of performance and usability. Firstly, it consistently identified a minimal yet highly informative set of features across all classifiers (SVM, RF, KNN), resulting in a compact model that does not compromise accuracy. On average, the hybrid method improved classification accuracy by 3–7% compared to traditional filter, wrapper, and embedded methods. Notably, the reduced feature set also led to substantial gains in computational efficiency, lowering inference time on edge hardware (Raspberry Pi 4 Model B) to under 0.5 seconds per instance, thus enabling real-time insect detection. Additionally, SHAP visualizations provided insight into feature relevance, revealing that the retained features–such as shape symmetry and size–were not only statistically significant but also biologically meaningful, corroborating domain knowledge from entomology. This interpretability strengthens the case for deploying the system in real-world agricultural settings where trust and transparency are crucial.

Comparison with existing studies

Figure 22 presents a comprehensive set of visual results showcasing the performance of the proposed insect classification system across multiple images. Each row demonstrates the original and predicted labels for various insect species, namely the Colorado potato beetle (CPB), ladybird, and aphid, within different host plant contexts (Solanum tuberosum, Solanum melongena, and Solanum lycopersicum).

Fig. 22

A comprehensive set of visual results showcasing the performance of the proposed insect classification system across multiple images.

Bounding boxes in various colors highlight the insects’ locations and help illustrate the classification decisions. For example, the images for CPB show red or yellow bounding boxes around the insect’s body, while the photos for aphids display pink bounding boxes. These visual comparisons highlight the model’s ability to accurately locate and classify the insects, even under challenging visual conditions like overlapping leaves or occlusions. In the images, it’s evident that the proposed approach is robust across varying image conditions, correctly predicting insect species even in cluttered scenes. The bounding boxes accurately capture the insects’ positions with minimal false positives. Furthermore, the figures indicate that the model maintains consistent performance across diverse plant backgrounds, reinforcing the system’s generalization ability. These results collectively underscore the feasibility of deploying this classification and localization system for precision agriculture, enabling targeted pest management interventions based on real-time insect detection and classification. To further evaluate our proposed method’s effectiveness, we compared it with existing studies, focusing on classification accuracy, as many authors do not report computational complexity. Using the widely known IP102 dataset for insect classification,³⁸ achieved about 1.48% higher accuracy than our RF and SVM models. This slight difference can be attributed to our approach of reducing redundant features, trading off a small amount of accuracy for faster computation. Furthermore, their work treats the entire image as a single insect species for classification, without considering object detection or insect localization. Interestingly, our RF model achieved 92.62% accuracy, outperforming the 91.5% reported by³⁹ using a CNN-based model that included all features, highlighting the strength of our feature selection approach. Likewise, our SVM model’s 91.16% accuracy exceeds previous studies’ results⁴⁰ at 85.2% and⁴¹ at 86.81%), even though their task was more straightforward (just distinguishing positive and negative samples) and their models were more computationally intensive. This clearly shows that incorporating feature selection boosts performance and reduces computational demands, enhancing generalizability by prioritizing the most relevant dataset features. Our study also considers the limitations of deep learning methods, which generally require extensive data and substantial computing power to produce reliable models (Cserni and Rovid, 2023). However, real-world insect-plant datasets are often scarce, making it hard to train robust deep learning models for practical applications. Many researchers compensate with data augmentation, which may not fully reflect real-world conditions. Visual results (as shown in Fig. 12) confirm that our method successfully pinpoints insect locations within various complex backgrounds, which is crucial for precision spraying or targeted pest removal. Though we used three different crop plants to improve data diversity, we did not evaluate the influence of individual host plants on model performance. Future work could address this and the challenge of detecting multiple insect species in the same image–overlapping objects could hinder detection accuracy. While deep learning models may offer marginally higher accuracy, our proposed method delivers competitive performance with significantly lower hardware demands–an advantage for agricultural robots or in-field devices where size, energy, and computational resources are constrained. It’s also worth noting that for agricultural pest management, a perfect detection rate is not always essential–an 80% accuracy can still effectively reduce pest numbers below the economic threshold³⁴. Our method consistently achieves or exceeds this level while requiring less computational power, making it cost-effective and feasible for practical, real-world agricultural applications.

Discussion

The proposed hybrid-XAI framework is inherently scalable for high-volume monitoring because inference is performed on a compact, hybrid-selected feature subset and lightweight classifiers, yielding low computational cost per image while preserving accuracy (Fig. 19). Since the model retains $\ge 90\%$ accuracy with only the top-11 features, it can support batch or streaming analysis from smart traps, field cameras, or UAV pipelines in large farming operations without requiring heavy cloud resources^25,33. Moreover, our dataset already includes multiple crop backgrounds and diverse field conditions, and the SHAP+PFI intersection favors stable, biologically meaningful features that generalize across crop-specific variability⁴. Nevertheless, we acknowledge that broader agro-ecological coverage (more crop types, regions, and seasonal contexts) will further strengthen scalability claims; thus, future work will expand multi-crop datasets and validate throughput in continuous sensing/UAV deployments^3,13,42.

The results of this study highlight the critical role of feature selection and hybrid modeling in achieving high-accuracy insect classification while maintaining computational feasibility for real-world applications. As summarized in Table 5, the hybrid model integrating SVM and GBM achieved an overall accuracy of 96.7%, with precision, recall, and F1-score values of 96.5%, 96.9%, and 96.7%, respectively. These results surpass the individual performance of the SVM (accuracy: 94.1%), RF (accuracy: 93.2%), and GBM (accuracy: 95.0%) models, underlining the effectiveness of the hybrid approach in leveraging complementary model strengths. Importantly, the proposed explainable artificial intelligence (XAI)-based feature selection method achieved superior accuracy while significantly reducing computational complexity compared to conventional feature selection techniques. For instance, the XAI method consistently outperformed methods such as mutual information (MI), the Fisher separation criterion, and the maximal information coefficient (MIC), achieving 90% accuracy with only the top 11 selected features (Fig. 11). This compares favorably to the MI method’s maximum accuracy of 88% and variance thresholding’s 85%, demonstrating that careful feature selection can mitigate the risk of overfitting while preserving robust classification performance. Notably, adding the first three histogram of oriented gradients (HOG) features in the XAI-based approach boosted accuracy to 65%, affirming HOG features’ strong discriminative power. However, excessive feature inclusion (beyond 11 features) led to fluctuations in accuracy across all tested methods, underscoring the importance of balancing informativeness and noise.

Further supporting this finding, Shapley values (Fig. 8) confirmed that features such as area length ratio (ALR), roundness, length perimeter ratio (LPR), compactness, and circularity had the greatest average impact on model performance, while features such as energy and rectangularity had minimal impact. These observations validate the XAI-based selection, which prioritizes features that consistently contribute to accurate classification. This also explains the hybrid model’s strong performance, as the selected features provide a comprehensive representation of insect shape, texture, and color, which are crucial for distinguishing species in complex field environments. Comparing our method with existing literature reveals competitive or superior performance. For instance,³⁸ achieved a slightly higher accuracy (about 1.48% more) than our RF and SVM models on the IP102 dataset. Still, their approach focused on full-image classification and did not address object localization. In contrast, our RF model achieved an accuracy of 92.62%, outperforming the 91.5% reported by Kasinathan et al.³⁹ using a CNN with all features. Similarly, our SVM model achieved 91.16% accuracy, surpassing the 85.2% and 86.81% reported by Yao et al.⁴⁰ and Liu et al.⁴¹, respectively. These improvements highlight the strength of integrating feature selection into the pipeline, as it enhances model generalizability while reducing computational burden.

The computational feasibility of our approach is equally noteworthy. The proposed method required approximately 4 hours and 13 milliseconds for training and prediction, respectively, with the SVM-based model, resulting in a model size of around 50 MB and an accuracy of 91.16%. These modest requirements make it suitable for deployment on edge devices such as the Raspberry Pi or lightweight cloud-based systems. Unlike deep learning models, which demand large datasets and considerable computational resources⁴³, our method is more adaptable to real-world agricultural scenarios, where data availability and hardware resources are often limited. Visual examples (Fig. 12) demonstrate the effectiveness of the proposed pipeline in accurately locating insects within cluttered natural backgrounds. This is critical for practical pest management applications, such as targeted pesticide spraying or precision biological control. While this study did not explicitly evaluate the impact of different host plants on model performance, it remains a promising area for future work. Additionally, scenarios involving multiple insect species in a single image could pose challenges due to overlapping objects, warranting further investigation.

Although the SHAP- and PFI-guided explanations were generally consistent with entomological knowledge by emphasizing biologically meaningful cues (e.g., CPB elytral stripe patterns and body contour, ladybird spot – roundness traits, and aphid colony texture), we observed a small number of challenging cases in which model explanations diverged from expert human interpretation. Such mismatches primarily occurred when insects were partially occluded by foliage, imaged from extreme angles, or captured under low-contrast or uneven illumination, leading the model to assign higher attributions to secondary cues, such as nearby leaf-edge gradients or localized texture patches, rather than the canonical insect body parts that experts typically prioritize. In densely infested aphid scenes, SHAP occasionally highlighted colony-level texture regions instead of individual insect boundaries, which remains useful for automated detection but differs from how specialists visually confirm species. These divergences likely reflect real-field visual complexity and background entanglement rather than systematic model bias. Expanding training data with stronger occlusion/lighting variability and integrating instance-level segmentation or attention constraints may further improve alignment between XAI outputs and expert reasoning in future work.

Despite the strong performance of the proposed hybrid SHAP+PFI framework, several limitations should be acknowledged. First, while dual-XAI selection improves stability and interpretability, SHAP attributions can become less reliable in challenging field scenarios such as severe occlusion, extreme insect poses, or low-contrast illumination, where background textures or leaf edges may receive non-trivial importance and partially diverge from expert reasoning^31,32,44. Second, although our dataset was collected under diverse real-farm lighting and background conditions, it covers three insect species and one negative class; therefore, generalization to a wider ecosystem containing many pest and beneficial species, or to multi-species images with overlapping insects, requires further validation on larger-scale benchmarks^5,33,45. Third, the edge-deployment claims are supported by latency and model-size measurements, but we did not yet profile memory usage, power consumption, or thermal stability during long-duration field operation, which are critical for confirming real-time reliability on low-cost devices^23,25. Finally, feature selection is performed offline; thus, the current pipeline does not support online adaptation or incremental updates when new insect species or unseen field conditions arise. Addressing these issues through broader datasets, instance-level segmentation, extended hardware benchmarking, and adaptive learning will further strengthen the system’s practical readiness.

Finally, it is essential to emphasize that agricultural pest management does not always require near-perfect detection rates. An accuracy of 80% or higher can reduce pest populations below the economic threshold³⁴, balancing the need for precision with practical deployment constraints. Our method meets or exceeds this threshold while maintaining low computational demands, making it a cost-effective and accessible solution for precision agriculture applications. By integrating explainable feature selection with hybrid classifiers, this study contributes to the development of intelligent pest control systems that are accurate, computationally efficient, and robust to environmental variability. The synergy between wrapper-based feature selection and hybrid modeling significantly enhances insect classification performance, mitigates overfitting, and ensures computational feasibility for deployment on resource-constrained hardware. These findings provide a robust foundation for advancing precision pest management strategies through explainable, efficient machine learning techniques.

Our findings reveal that the hybrid XAI-based feature selection method not only improves classification accuracy but also significantly reduces model size and inference time. This enables real-time detection on hardware-constrained platforms. Additionally, the explainability of the selected features enhances trust and potential adoption among end users, such as farmers and agronomists. The interpretability provided by SHAP and PFI empowers users to understand why certain insects are detected as pests or beneficial, aligning with the goals of responsible AI in agriculture. Future work could explore extensions of online learning and its integration with UAVs or edge sensors for broader, scalable pest monitoring.

Conclusion

This research addresses a critical gap in integrating explainable feature selection with hybrid machine learning models for insect classification, with particular emphasis on balancing predictive accuracy and computational feasibility for real-world agricultural applications. While many prior studies have prioritized classifier accuracy or purely deep learning architectures—often requiring substantial labeled data and high computational resources—this study introduces a dual-XAI hybrid feature selection strategy using SHAP and Permutation Feature Importance (PFI), paired with efficient classifiers (SVM and GBM), to enhance both interpretability and performance under resource constraints. Experimental results on a curated, balanced field dataset demonstrate strong empirical performance, achieving an overall accuracy of 96.7%, with precision, recall, and F1-score values of 96.5%, 96.9%, and 96.7%, respectively. Importantly, competitive accuracy (90%) is retained using only the top 11 hybrid-selected features, confirming that the proposed SHAP+PFI intersection effectively reduces redundancy while preserving discriminative power. This stability surpasses conventional baselines such as mutual information and variance thresholding, which reach lower peak accuracies of 88% and 85%, respectively. From a deployment perspective, the trained model exhibits lightweight characteristics suitable for low-cost hardware, with an approximate size of 50 MB and prediction latency of 13 ms per sample on a Raspberry Pi 4 Model B. However, we acknowledge that these metrics alone do not fully establish edge compatibility; memory utilization, energy consumption, and thermal behavior under prolonged field operation were not measured in the present study and remain essential benchmarks for future validation. Overall, the key contribution of this work lies not only in improved accuracy but also in increased transparency and compactness, which together support greater trust and operational viability in smart agriculture. The SHAP-based explanations further confirm that retained features align with biologically meaningful insect traits (e.g., size, symmetry, and texture cues), strengthening both model reliability and interpretability. Future work will extend this framework by:

• profiling energy and memory performance during extended field trials;

• integrating deeper feature embeddings (e.g., CNN-based representations) within the same dual-XAI selection paradigm;

• evaluating generalization across broader crop backgrounds and additional insect species, including multi-species scenes;

• deploying real-time monitoring pipelines in live precision farming and IPM workflows.

These directions will further solidify the practical readiness of the proposed system and advance robust, explainable AI solutions for sustainable pest and beneficial insect management. In practical deployment, the proposed hybrid-XAI model can be integrated into smartphone-based scouting or low-cost edge-trap systems, where users capture leaf images and receive real-time pest/beneficial insect classification, with SHAP-based visual justification to support trustworthy IPM decisions. Since the model preserves high accuracy with only 11 features and low inference latency on edge hardware, it is suitable for on-device or near-edge monitoring applications that reduce unnecessary pesticide use while enabling timely field interventions.