Medicinal plant leaf disease classification using optimal weighted features with dilated adaptive DenseNet and attention mechanism

Abstract

The agriculture sector plays a pivotal role in the growth of the global economy, but remains highly susceptible to prediction errors, particularly in disease identification. To address the limitations of existing approaches, this study proposes a deep learning-based framework for the classification of medicinal plant leaf diseases. “Medicinal plant leaf images are collected from the standard data source. These images undergo a pre-processing phase that includes filtering and Contrast-Limited Adaptive Histogram Equalization (CLAHE) to enhance visual quality. Subsequently, an adaptive thresholding mechanism is employed for precise leaf segmentation. For effective disease recognition, deep feature extraction is carried out using a customized Multi-Scale VGG16 architecture,” capturing diverse features such as color, shape, and texture. These heterogeneous features are then subjected to a weighted fused feature selection process, where feature weights are optimized using a novel Hybridized Zebra with Krill Herd Optimization (HZKHO) algorithm. The optimized feature set is input to the disease classification stage, which employs an Attention-based Dilated Adaptive DenseNet (A-DADensenet) model to produce accurate classification results. The proposed model achieves an impressive classification accuracy of 90.69%, thereby demonstrating its effectiveness in accurately identifying diseased medicinal plant leaves. “The integration of deep learning with a hybrid optimization technique” significantly enhances the model’s classification performance, proving its potential for real-world agricultural applications.

Introduction

Herbs have been a significant part of society from the beginning of human civilization, and specifically, they improve the well-being and fitness of humans. Rather than conventional medication, many people prefer using herbal balms for their health benefits¹. Hence, herbs are assumed as the safest option, and they contain disease prevention as well as health-promoting properties of phytochemicals. Many medicinal plants and herbs provide substantial benefits to human health². A huge number of herbs have been discovered in the current era, but still, some of them are unidentified and undiscovered³. Therefore, the assessment of medicinal plants is significant because most of these plants are mainly utilized for the production of topical products and human food. On the other hand, medicinal plants help regulate and maintain a balance of atmospheric gases, including oxygen, carbon dioxide, and nitrogen compounds, and the plants give some nutrients to humans⁴.

The health condition of the medicinal plants is determined via the real-time recording and monitoring systems, and it provides significant aspects to protect the lives of the medicinal plants⁵. Most of the plant diseases arise from viral, bacterial, and fungal infections that may affect the entire growth rate. Therefore, early-stage detection of plant diseases helps maximize plant growth, which in turn contributes to the nation’s agricultural development. The plant protection is manually done by the farmers with the help of their own experience and expertise⁶. In this process, farmers spend considerable time and effort visiting their farms to record and manually inspect the condition of the plants⁷. Manual methods for detecting and identifying plant species and diseases require high energy consumption and are highly prone to human error. The determination of plant diseases at the initial phase is very difficult and challenging⁸. Moreover, the recognition and early detection of herbal disease types could be achieved via analysis of venation architecture, texture, color, and shape. The early symptoms of leaf diseases can be identified from the changes in the leaf’s color, while the changes in texture, venation patterns, and shape are also significant aspects for classifying disease types⁹. Researchers have developed various automated approaches for detecting leaf diseases based on these visual features¹⁰. Different variants of plant leaf diseases can be detected and classified through structured phase-wise analysis, supported by effective pre-processing methodologies. Moreover, the accuracy of the plant leaf disease classification process can be enhanced by using machine learning approaches, which reflects the performance of recognition and detection of species and diseases at an early stage with high accuracy¹¹.

Microbiological studies are performed based on molecular approaches and the implementation of pathogenic clusters in immunological laboratories¹². However, implementing these approaches in real-world laboratory environments is challenging¹³. Since the microscopic and manual evaluations are not highly effective and automated, and hence the automatic identification of plant leaf diseases is highly desirable in the area of diagnostics¹⁴. To address these limitations, advanced deep learning and machine learning methodologies have been developed to improve the detection of plant leaf diseases¹⁵. A series of image processing steps is typically applied, which includes steps such as the acquisition of images, image pre-processing for quality enhancement, segmentation of images, extraction of features, and classification¹⁶. Many of the traditional approaches for the identification of diseases utilize hand-crafted features that include shape, SIFT, color, and HoG. These features provide the most discriminating and useful information from the images that are subjected to the classifier¹⁷. However, these techniques still struggle to learn the complex distribution of the images that differentiate various classes of diseases¹⁸

Rationale of this study

The prior frameworks can effectively detect the medicinal plant leaf disease, but they suffer from several crucial issues like error, optical illusion, and high computational time¹⁹. The traditional approach has suffered from the bottleneck effect. Moreover, traditional networks cannot learn and retrieve the features directly from the high-dimensional dataset²⁰. Traditional methods of leaf disease diagnosis rely heavily on human expertise and visual inspection, which can be time-consuming, labor-intensive, and prone to errors²¹. The problems in the classical methods can be effectively overcome by the “deep learning-based disease detection models,” which help to enhance the performance in the medicinal plant leaf disease classification. So, this work developed an advanced deep learning-based method for medicinal plant leaf disease classification. The study focuses on the development of a novel A-DADenseNet model that integrates attention mechanisms, dilated convolutions, and adaptive dense connections. This architecture is designed to capture the complex and variable patterns present in leaf disease images, and to enable automatic feature extraction and classification. The incorporation of HZKHO optimization is also a key aspect of this study, as it enables the adaptive adjustment of hyperparameters to optimize model performance. “By leveraging the advancements in deep learning techniques” and addressing the limitations of existing models, this study aims to make a significant contribution to the field of plant disease diagnosis and management.

Novelty of the proposed work in comparison with the latest literature published in 2023 and 2024

The research proposes a novel framework for medicinal plant leaf disease classification, distinguishing itself from existing methods through its unique combination of techniques. Recent studies explored techniques such as ensemble-based deep learning models²², hybrid deep learning models²³ and lightweight CNN²⁴ for the classification of the medical leaf plants. In contrast, "the novelty of the novel approach” lies in its integration of a hybrid optimization algorithm (HZKHO) with a deep learning model (A-DADensenet). This signifies a focus on optimizing the deep learning model’s performance through a specialized optimization algorithm, rather than solely relying on architectural variations or ensemble methods. The core innovation lies in the synergistic combination of an optimization algorithm with a “deep learning model” for enhanced classification of medical leaf plants. Furthermore, traditional methods often rely on simpler feature extraction techniques, which may not capture complex patterns in plant leaves. The proposed framework aims to overcome these limitations by leveraging the power of Multi-Scale VGG16 feature extraction, capturing a wider range of features. In addition to this, the proposed model is capable of identifying various diseases in medicinal plants. This new approach aims to contribute to the field by proposing a novel framework that achieves high accuracy in medicinal plant leaf disease classification, demonstrating its potential for real-world applications. Moreover, many existing models rely on manual feature extraction, which can be time-consuming and may not capture all relevant features. In contrast, the A-DADenseNet model’s attention mechanism enables automatic feature extraction, reducing the need for manual intervention.

The contributions of the proposed model are summarized below.

• To implement a new “classification of medicinal plant leaf disease method” using an advanced deep learning network to recognize and classify the diseases at the starting stage, which may help to make preventive measures based on the classified diseases.

• To design an HZKHO for optimizing the weights to get fused features, and it selects optimal features from the retrieved features. Moreover, the parameters from DenseNet are tuned via HZKHO to maximize the classification performance.

• To get weighted fused features from the extracted features using HZKHO, and hence, the best features are obtained without losing any relevant information.

• To design an A-DADensenet for classifying the medicinal plant leaves to identify the type of diseases to be affected, where the hidden neuron count, epochs, and activation function are optimized by using proposed HZKHO to enhance the classification score by the false positive rate, negative predictive value, and precision.

The remaining portions used in the investigated HZKHO-A-DADensenet-based medicinal plant leaf disease classification model are summarized as follows. Existing works with their merits and disadvantages are illustrated in Section II. Architectural representation and dataset description are given in Section III. The segmentation process and the feature extraction are given in Section IV. The classifications model and developed HZKHO are explained in Section V. Results are discussed in Section VI, and the conclusion is illustrated in Section VI.

Literature survey

Related works

In 2020, Mustafa et al.²⁵ suggested an early disease diagnosis methodology using machine learning for disease classification in the herb species via the support of computer vision, where the color, odor, texture, and shape of the features have been retrieved from the images using a hybrid intelligent system. The hybrid intelligent system demonstrated high recognition accuracy and detection rate, which might help in detecting the disease at first.

In 2019, Geetharamani and Pandian²⁶ have investigated a classification framework for diseases from plants using a Deep Convolutional Neural Network (Deep CNN) in the background. Data augmentation methods such as gamma correction, color augmentation, Principal Component Analysis (PCA), image flipping, scaling, rotation, and noise injection were used.The DCNN model was trained with network parameters including epochs, dropout, and batch size. The simulation results showed that the DCNN-based framework achieved higher recognition accuracy than previous machine learning methods. Moreover, the effectiveness has been validated through reliability and consistency.

In 2022, Sathiya et al.²⁷ have initiated an automatic early disease detection and classification technique for herbal plants with the utilization of hybrid soft computing approaches. Leaf segmentation was carried out using a multi-swarm-based Coyote Optimization (CO) algorithm to isolate the diseased region. Finally, classification was performed using a fitness-distance balanced Deep Neural Network (DNN) to ensure efficient classification results.

In 2022, Mustafa et al.²⁸ have recommended a framework using a five-layered CNN network for determining various types of diseases from the plant leaves, which trained the network using 20,000 augmented images from the traditional benchmark datasets. The proposed framework effectively detected pepper bell plant diseases using the five-layered CNN. The analytical outcome demonstrated that the optimized CNN approach achieved high accuracy. Its robustness was significantly improved under different cultivation environments.

In 2021, Ahmad et al.²⁹ have offered a CNN-based system to classify plant disease symptoms systematically with high efficiency. The system was highly memory-efficient and supported faster deployment in industrial applications through reduced training times. To address class imbalance caused by uneven sample distribution, a simple statistical approach was applied. To avoid negative transfer effects, a stepwise transfer learning approach was implemented, ensuring a high convergence rate and minimizing overfitting. Using the PlantVillage dataset, the proposed CNN-based system outperformed existing models and achieved higher classification accuracy.

In 2022, Patle et al.³⁰ have suggested an early disease classification approach based on the information on Leaf Wetness Duration (LWD). Leaf Wetness Sensors (LWS) were developed on flexible polyamide substrates using “Graphene Oxide (GO) as a sensing film to detect water molecules on the leaf canopy. The fabricated sensors were tested under various lab conditions, and their performance in sensing water molecules and air was analyzed. Test results showed that the proposed LWS model provided a significant improvement in LWD value compared to commercial sensors.

In 2023, Sahu and Pandey³¹ have presented a plant foliar disease detection approach using the hybridization of a Random Forest with a Multiclass Support Vector Machine (HRF-MCSVM) network to provide extensive disease section results. Image features were first preprocessed and segmented using the Spatial Fuzzy C-Means method. Final classification was performed using the HRF-MCSVM model, “which demonstrated superior performance compared to existing methods.”

In 2021, Muthukrishnan et al.³² have developed a plant leaf disease detection scheme to recognize the diseases in plant leaves. Spherical polar coordinates were used to identify diseases based on Hue, Saturation, and Intensity (HSI) values. Philodendron leaves were processed, and specks were detected in the leaves. The technique involved converting natural color images to grayscale before applying the HSI model. Double-sized images were used to enhance the visibility of affected regions, aiding spot identification. Detected specks were highlighted using a brighter intensity scale.

In 2021, Azadnia et al.³³ have proposed a robust image-processing algorithm for the accurate identification of medicinal plants. Shape, color, and texture features were extracted from the collected images. Artificial Neural Networks (ANNs) were applied to classify the plants and evaluate the error rate, accuracy, and correlation. The model achieved 100% accuracy, demonstrating its reliability in classifying medicinal plants effectively.

Problem statement

The medicinal plant leaf disease detection method is evaluated based on the symptoms presented in various parts of the plants. However, it requires extensive knowledge to identify the disease accurately; otherwise, it can negatively impact the production outcome. The advantages and limitations of conventional approaches related to medicinal plant leaf disease are displayed in Table 1. PNN and SVM²⁵ have better accuracy and reliability in detecting the disease at an early stage, and they are simple to design and implement. However, they require significant computational time to analyze the data during the testing phase. Deep CNN²⁶ has a higher reliability and consistency rate than existing models; however, it demands intensive training, especially in the absence of labeled images. FDB-DNN²⁷ offers a higher disease prediction rate, selects the optimal solution more effectively than traditional models, and efficiently minimizes the dimensionality problem. But, “it requires a large amount of computational time to process the data.” CNN²⁸ has high robustness and effectively minimizes training time, but it requires more training data and suffers from overfitting issues. CNN²⁹ has higher memory efficiency, reduces computational time, and effectively lowers convergence complexity and negative learning. However, it still faces gradient exploitation issues and does not encode the object’s positional information. GO³⁰ has lower intrinsic toxicity than other models, but it is expensive, and the task is considered complex. HRF-MCSVM³¹ enhances computational accuracy and has a greater memory capacity for storing data. However, it operates slowly and demands high computational power. Histogram Equalization³⁴ uses a huge factor to produce accurate results in identifying plants and utilizes a bot to trace plant images from different angles for better analysis. However, it does not support precise detection and depends largely on maximum and minimum variable values. Thus, to overcome the aforementioned limitations in conventional approaches, there is a critical need to develop an innovative framework using an advanced deep learning algorithm.

Table 1 Features and challenges of existing medicinal plant leaf disease classification models.

Motivation

“Medicinal plant leaf disease determination is a significant research area in the field of machine vision.” Machine vision-based approaches address the challenges encountered in conventional naked-eye-based medicinal plant leaf disease detection models. Uniformly illuminated images are obtained by choosing the appropriate light source and shooting angle, and these images help improve the efficiency and accuracy of medicinal plant leaf disease classification. Such techniques reduce the complexities involved in designing classical algorithms but tend to increase overall implementation costs³³. Moreover, the complete elimination of scene variation effects in classical algorithms under natural environmental conditions is unrealistic³⁵. As a result, disease detection approaches for medicinal plants face several issues, including low contrast, image noise in lesion regions, minimal variation between background and lesion areas, and significant variability within lesion regions³⁶. Furthermore, various disturbances occur during the collection of medicinal plant leaves under diverse environmental and lighting conditions. The limitations observed in classical methods are effectively addressed by deep learning-based disease detection models for medicinal plant leaves^37,38. Deep learning networks with multiple layers exhibit strong feature extraction and autonomous learning capabilities. Hence, to provide enhanced feature representation and improve the accuracy of medicinal plant disease detection, a novel deep learning framework is suggested.

Newly developed advanced deep learning-based model for the classification of medicinal plant leaf diseases with parameter optimization

Architectural view of the developed framework

The structural description of the developed medicinal plant leaf disease categorization model using advanced deep algorithms is given in Fig. 1.

Fig. 1

Structural representation of the implemented medicinal plant leaf disease image classification and segmentation approach.

A new advanced “deep learning-based disease classification method” from medicinal plant leaves is implemented to classify the types of medicinal plant leaf diseases with higher precision and identify the diseases at a very early stage. Initially, the required images are acquired from the mendely dataset, and then it is subjected to image preprocessing. Here, Median filtering and CLAHE are utilized to preprocess the images. This step helps remove noise and enhance the details in the images. Then, the preprocessed image is applied to adaptive thresholding-based leaf segmentation, where the threshold value is optimally selected using the newly recommended HZKHO. The segmented images are then used to extract deep features using Multi-Scale VGG16. Additionally, color, shape, and texture features are extracted from the segmented leaf images. The extracted features are then subjected to the weighted fused feature selection process. Here, the weights are optimized using the proposed HZKHO, and based on the optimized weights, the best features are selected. After, the selected fused features are given to the A-DADensenet to classify the diseases from the medicinal plant leaves with high classification accuracy. Here, the parameters like activation function, hidden neuron count, and epochs are optimized with the help of the suggested HZKHO. At last, the results are evaluated to showcase the performance of the developed classification of medicinal plant leaf disease scheme.

Collection of plant leaf images

The description of the dataset utilized for the medicinal plant leaf disease is given as follows.

The medicinal plant leaf images are garnered from the dataset of Mendeley Data that is available in the data source of https://data.mendeley.com/datasets/hjrhrt5hs8/2 with access date: 2025-07-21. Soursop is one of the herbal plants, and it is mostly used as an anticancer, antidiabetic, insecticidal, anti-inflammatory, antibacterial, antimalarial, and antioxidant plant. The images of medicinal plant leaf diseases are included in this dataset. Image acquisition has been taken in different places, such as Bangladesh, Kushtia, and Dhaka. This dataset contains six diseased classes, including Cutting Weevil, Cutting Caterpillar, Die Back, White Fly, Healthy, and Yellow. The original size of the images is 3024 × 4032 pixels, which is compressed into 1024 × 1024 pixels to reduce the memory storage size of the data. Totally, 3838 images are presented in this dataset, and the format of the data is JPG. The normal and the affected medicinal plant leaf images are shown in Fig. 2.

Fig. 2

Collected sample medicinal plant leaf images.

Various traditional databases are used to gather the required medicinal plant leaves for the classification of various diseases indicated by the term $MP_{f}^{in}$, where the term $f = 1,2,3,...,F$. From this, $F$ it gives the total count of image samples.

Image preprocessing

The collected images are passed through an image preprocessing stage, where the image data is improved by suppressing or eliminating unwanted distortions and enhancing the relevant features required for subsequent classification. Here, preprocessing techniques such as median filtering and CLAHE are utilized for enhancing the quality of the image data. A detailed explanation of median filtering and CLAHE is provided below.

Median filtering

The collected image $MP_{f}^{in}$ is provided as the initial input. It is a non-linear filtering technique commonly used to remove unwanted noise from input images. The main benefit of using median filtering is that it preserves image edges while reducing noise. It functions by sliding a window over the image pixel-by-pixel, and then replacing the central pixel value with the median value of its neighboring pixels. The set of neighboring pixels is referred to as the window. The window slides across the image pixel-by-pixel. All pixel values within the window are arranged in numerical order, and the center pixel is replaced with the middle value. The result of this process is referred to as the filtered output $MP_{f}^{Lt}$.

CLAHE

The median-filtered image is provided as input to CLAHE, which is denoted by $MP_{f}^{Lt}$. CLAHE is used to enhance contrast and improve differentiation between medicinal leaf images. In this method, the histogram of the image is divided into several predefined sections. Then, the image intensity is adjusted and redistributed uniformly across the grayscale range. The enhancement function is applied to all neighboring pixels, and based on the resulting pixel values, a transformation function is applied. CLAHE can be applied to both color and grayscale images. A clip limit is introduced to restrict the amplification of noise in homogeneous regions. The key input parameters obtained from the original image include the number of regions in the row and column, the clip limit, the distribution parameter type, and the dynamic range. The image is then partitioned into contextual regions, and the enhancement process is applied to each region. Next, a clipped histogram is created and mapped to gray levels. Finally, the enhanced image is generated by interpolating the gray level mappings of neighboring regions. The resulting image after applying CLAHE is denoted by $MP_{f}^{CL}$.

Plant leaf image segmentation and weighted fused feature selection using hybridized zebra with krill herd optimization

Adaptive thresholding-based segmentation

Image segmentation is a process, generally utilized in digital image processing, to divide the image into multiple regions based on the characteristics of pixels in an image^39,40. By using this segmentation approach, the objects and the boundaries are effectively located, including lines, curves, and so on. In this developed model, adaptive thresholding is performed to enhance the efficacy of the final classification.

Adaptive thresholding⁴¹

The input images given to the adaptive thresholding process are denoted by $MP_{f}^{CL}$. A fixed threshold value does not perform well under poor illumination conditions that vary from one medicinal plant leaf image to another. In the adaptive thresholding method, grey-level information is combined with gradient information. This mapping provides better segmentation results. The optimal threshold value differentiates the darker parts from the images. Initially, the image consists of $a \times b$ pixels with $W$ gray levels. The spatial location of the pixel is denoted as $\left( {m,n} \right)$ that is supposed to be the gray level value $g\left( {m,n} \right)$. The mapping function is defined in Eq. (1).

$$g:a \times b \to Q$$

(1)

They perform smoothing by applying the image to the smoothing filter $F$ with a window size $d \times d$. The evaluation of the smoothed image is represented in Eq. (2).

$$g`\left( {m,n} \right) = \frac{1}{{d^{2} }}\sum\limits_{{\left( {i,j} \right) \in F}}^{{}} {g\left( {m + i,n + j} \right)} \forall \left( {m,n} \right) \in a \times b$$

(2)

After mapping, the total variance and the between-class variance were also computed. The variance value depends on the mean value of the pixels.

“Based on the mean and variance determination, the cumulative distribution function” is calculated to perform the thresholding. The stopping condition of the thresholding depends on the cumulative distribution function. The steps to be followed in the Adaptive thresholding process are given in Fig. 3. Finally, the segmented regions of the medicinal plant leaves are obtained as indicated by $MP_{t}^{AdT}$.

Fig. 3

Adaptive thresholding-based medicinal plant leaf segmentation.

Feature extraction steps

The segmented images $MP_{t}^{AdT}$ are passed to the feature extraction process. By using the Multi-Scale VGG16, and “the other features such as color, shape, and texture are separately retrieved from the segmented images.”

Deep feature extraction using multi-scale VGG16⁴²

In the general VGG16, the filters use the same receptive field, and hence there is a chance for loss of scale-invariant features. In the multi-scale VGG16, a multi-scale block is added after the four max-pooling layers in the VGG16 network. Each multi-scale block has three convolutional layers with different receptive field scales. The first layer’s receptive field is considered as $3 \times 3$, and the other layers are assumed as $1 \times 1$. By using these various receptive fields, the multi-scale block captures more relevant features. The output sizes of each multi-scale block are the same, and diverse features are extracted using the output stride value. After the fifth max-pooling layer, all the features obtained from the multi-scale blocks are concatenated⁴³. “This concatenated output is then passed through the three fully connected layers in the network.” The final features are obtained from the softmax layer. The retrieved deep features are denoted by the term $Df_{z}^{VGG}$. The functional representation of the multi-scale VGG16 for deep feature extraction is given in Fig. 4.

Fig. 4

Deep feature extraction using multi-scale VGG16.

Color feature extraction

The color space represents color in terms of intensity values. Before extracting color features, it is important to define, visualize, and generate color using a color space approach. “The most commonly used color feature extraction methods include histogram intersection, Zernike chromaticity distribution, and color histograms”.

Histogram intersection

This method considers global color features. Its effectiveness mainly depends on the number of bins, although increasing the number of bins can also increase computational complexity.

Color histogram

This method represents the image from multiple perspectives. Colors are expressed as a frequency distribution, and the color frequency within the image is analyzed. Using the histogram values, issues related to translation, rotation, and scaling can be addressed.

Zernike chromaticity distribution

In this method, color features are extracted from the chromaticity space, providing fixed-length feature vectors that represent color characteristics. However, under flipping and rotation, the color content of the image may vary, affecting feature consistency.

In the developed method, the color feature extraction uses the color histogram approach. It is represented by $CL_{a}^{fs}$.

Shape feature extraction

“Shape features play a crucial role in object recognition. Accurately determining the location of these features can be challenging”. In this approach, a binary image algorithm is employed to extract shape features. Initially, the image is converted into a binary format consisting of black and white regions. Then, the outer boundary regions are precisely traced. The shape factor, used to quantify shape characteristics, is determined using Eq. (3).

$$shape\,factor = \frac{Total\,area}{{\left( {Diameter} \right)^{2} }}$$

(3)

Here, all the extracted boundaries and objects are labeled, and then shape features are extracted. The obtained shape features are indicated by $SP_{b}^{fl}$.

Texture feature extraction

The Gray Level Co-occurrence Matrix (GLCM) is used to extract texture features. The elements in the matrix represent the relative frequency of pixel intensity pairs, and by computing statistical measures, the texture features are extracted. The expression for evaluating texture features is given in Eq. (4).

$$TX_{c}^{fx} = \sum\limits_{m = 1}^{P} {\sum\limits_{n = 1}^{Q} {\left\{ {\begin{array}{*{20}c} {1,} & \begin{gathered} if\,J\left( {m,n} \right) = j\,and\, \hfill \\ J\left( {m + g_{m} ,n + g_{n} } \right) = j \hfill \\ \end{gathered} \\ {0,} & {otherwise} \\ \end{array} } \right.} }$$

(4)

The obtained features from the images are denoted by the term $TX_{c}^{fx}$. The extracted deep features $Df_{z}^{VGG}$, color features $CL_{a}^{fs}$, shape features $SP_{b}^{fl}$_, and texture features $TX_{c}^{fx}$ are fed to further processing.

Proposed HZKHO

The developed HZKHO is utilized in the newly developed advanced deep learning-based disease classification system from medicinal plant leaves to improve the classification outcome.

Purpose

The weights are optimized using HZKHO and the optimal deep features—color, texture, and shape are selected with the help of the designed HZKHO.

Here, the Zebra Optimization Algorithm is adopted because it provides a simple solution during optimization, and the Krill Herd Algorithm offers better fitness convergence in the search space. However, the convergence rate of these algorithms is relatively low.

Novelty

Hence, an improved optimization algorithm is required in the proposed method to enhance classification performance. In the developed HZKHO, the new position is derived by combining the outputs of ZOA and KHO. Position 1 is generated by the Zebra Optimization Algorithm (ZOA), denoted by $C_{{_{h,t} }}^{{New,R_{2} }}$, and Position 2 is obtained from the Krill Herd Optimization (KHO), represented by $C_{k}^{New}$. The average of these two estimated positions yields the updated position formula in HZKHO. The updated position is expressed in Eq. (5).

$$C_{New} = mean\left( {C1,C2} \right) + \frac{1}{C1/C2}$$

(5)

Here, the term $C1$ is updated using ZOA and $C2$ is updated using KHO, and the term $C$ denotes the newly updated position. Based on the updated position, the best features are selected, and the best parameters are optimized in the developed model with a high convergence rate.

ZOA⁴⁴: It is generally a nature-inspired algorithm, and it is inspired by the behavior of zebras. Zebras are famous for their white and black striped coats on their bodies. Here, the stripes are vertically placed on its neck and body. It is an inhibitory agent for biting flies, and it is efficient in hiding zebras from predators. To formulate ZOA, behaviors like foraging and defense are significant.

It moves to the forage by opening the way from the pioneer zebra to other zebra in the foraging process. Under the guidance of the pioneer zebra, other zebras are moved from the herd to the plain.

In the initialization phase, the zebras are considered as members of their population. The zebras are considered the candidate solution, and the zebra in the solution space is used to discover the decision variables. The population is denoted by $C$, $C_{h}$ represents the $h^{th}$ zebra, and $C_{h,t}$ is the value for $t^{th}$ the problem. The total count of decision variables is denoted by $x$ and the total population of zebras is given by $y$.

In the foraging phase, the members of the population are upgraded when searching for forage based on the zebra behavior simulations. This foraging phase is mathematically modeled in Eqs. (6) and (7).

$$c_{{_{h,t} }}^{{New,R_{1} }} = c_{h,t} + \phi \cdot \left( {R_{h} - W \cdot c_{h,t} } \right)$$

(6)

$$c_{{_{h,t} }}^{{New,R_{1} }} = \left\{ {\begin{array}{*{20}c} {C_{{_{h} }}^{{New,R_{1} }} ,} & {O_{{_{h} }}^{{New,R_{1} }} < O_{h} } \\ {C_{h} ,} & {elsewhere} \\ \end{array} } \right.$$

(7)

The new status of $h^{th}$ the zebra is indicated by $C_{{_{h} }}^{{New,R_{1} }}$, the $t^{th}$ dimensional value is indicated by $c_{{_{h,t} }}^{{New,R_{1} }}$, the objective function is defined as $O_{{_{h} }}^{{New,R_{1} }}$ the best member is referred to as the pioneer zebra that is indicated by $R_{h}$, $\phi$ is the random attribute in the interval of $\left[ {0,1} \right]$, $W$ indicates the round factor, which is selected at $W = rnd\,\,\,\,\left( {1 + \phi } \right)$.

In the defense strategy, the population members’ positions are updated, where the predator attacks are described. If a lion attacks the zebra, it selects the escape strategy, and if the other predators attack the zebra, it selects the offensive strategy. The newly updated position in the escape strategy is explained in Eq. (8).

$$c_{{_{h,t} }}^{{New,R_{2} }} = \left\{ {\begin{array}{*{20}c} \begin{gathered} M_{1} :c_{h,t} + D.\left( {2\phi - 1} \right) \hfill \\ .\left( {1 - \frac{q}{{M_{it} }}} \right) \cdot c_{h,t} , \hfill \\ \end{gathered} & {\Pr ob_{h} \le 0.5} \\ {M_{2} :c_{h,t} + \phi .\left( {G_{h} - W.c_{h,t} } \right),} & {elsewhere} \\ \end{array} } \right.$$

(8)

Based on the objective function, the position is updated, and it is formulated in Eq. (9).

$$c_{{_{h,t} }}^{{New,R_{2} }} = \left\{ {\begin{array}{*{20}c} {C_{{_{h} }}^{{New,R_{2} }} ,} & {O_{{_{h} }}^{{New,R_{2} }} < O_{h} } \\ {C_{h} ,} & {elsewhere} \\ \end{array} } \right.$$

(9)

The modes of the strategy are indicated by $C_{{_{h} }}^{{New,R_{2} }}$, is the new status of $h\,\,^{th}$ zebra, $t\,^{th}$ dimensional value is indicated by $c_{{_{h,t} }}^{{New,R_{2} }}$, the objective function is defined as $O_{{_{h} }}^{{New,R_{2} }}$ the iteration contour is defined by $q$, $D$ is the constant number taken at $0.01$, $M_{it}$ is the maximum iteration, and $\Pr ob_{h}$ is the probability of selecting methods. Moreover, the status of the attacked zebras is defined by $G_{h}$. In this way, the solution is updated in ZOA.

KHO⁴⁵: It is formulated based on the krill swarms’ herding behavior in response to environmental and biological processes. It is generally a bio-based swarm intelligence algorithm, and it is formulated based on mechanisms like enhanced reproduction, environmental conditions, feeding ability, and protection from predators.

The individuals are removed in predation, which leads to a reduction in the distance and krill density, where the krill swarm from the food at a particular location. The three fitness functions to be carried out there are induced movement by the krill individuals, the foraging mechanism, and random diffusion.

Initially, the Lagrangian model is generalized in $v$ dimensional space as defined in Eq. (10).

$$\frac{{dD_{k} }}{dm} = C_{k} + G_{k} + F_{k}$$

(10)

The developed motion by the other krill individuals is denoted by $C_{k}$,$G_{k}$ is physical diffusion and $F_{k}$ is the foraging movement of $k^{th}$ krill individuals.

Induced motion by other krill individuals

The induced motion direction is calculated from the target, local and repulsive swarm density. The krill individuals try to maintain high density, and that is expressed in Eq. (11).

$$C_{k}^{New} = C^{Mx} \varpi_{k} + \psi_{p} C_{k}^{Old}$$

(11)

The inertia weight is denoted by $\psi_{p}$, $C\,^{Mx}$ is the speed,$C_{k}^{Old}$ the last induced motion, and $\varpi_{k}$ the induced motion and it is estimated from Eq. (12).

$$\varpi_{k} = \varpi_{k}^{Lo} + \varpi_{k}^{Tt}$$

(12)

The neighbor krill provides the local effect, and it is denoted as $\varpi_{k}^{Lo}$, and the best krill provides the target direction effect, which is indicated by $\varpi_{k}^{Tt}$.

There is a repulsive or attractive tendency between the individual, and hence the effect of neighbors for a local search needs to be determined as given in Eq. (13).

$$\varpi_{k}^{Lo} = \sum\limits_{m = 1}^{SS} {\hat{I}_{k,m} } \hat{D}_{k,m}$$

(13)

$$\hat{D}_{k,m} = \frac{{D_{m} - D_{k} }}{{\left\| {D_{m} - D_{k} } \right\| + \varepsilon }}$$

(14)

$$\hat{I}_{k,m} = \frac{{I_{k} - I_{m} }}{{I^{Wst} - I^{Bst} }}$$

(15)

“The worst and best fitness values of the krill are denoted by $I^{Wst}$ and $I^{Bst}$, respectively”. The fitness function value of the $k^{th}$ krill is represented by $I_{k}$, and $m^{th}$ neighbor fitness is indicated by $I_{m}$, where $m = 1,2,...,SS$. Here, the term $SS$ denotes the number of neighbors. The distance is given in Eq. (16).

$$Ed_{q,k} = \frac{1}{5S}\sum\limits_{m = 1}^{S} {\left\| {D_{k} - D_{m} } \right\|}$$

(16)

The number of krill individuals is indicated by $S$ and $Ed_{q,k}$ is the sensing distance. The fitness function is defined in Eq. (17).

$$\varpi_{k}^{Tt} = Q^{Bst} \hat{I}_{k,Bst} \hat{C}_{k,Bst}$$

(17)

The effective coefficient of the $k^{th}$ krill individual is termed as $Q^{Bst}$, and the solution to the global optima is led by the factor $\varpi_{k}^{Tt}$. The effective coefficient $Q^{Bst}$ is computed using Eq. (18).

$$Q^{Bst} = 2\left( {\phi + \frac{It}{{M_{It} }}} \right)$$

(18)

“The term $It$ is the original iteration count and $M_{It}$ is the maximum iteration count.” Moreover, the term $\phi$ is the random attribute that lies in the interval of $\left[ {0,1} \right]$.

Foraging movement

The food location is effectively described in the foraging movement. This foraging movement is described in Eq. (19).

$$G_{k} = Z_{g} \zeta_{k} + \psi_{g} G_{k}^{Old}$$

(19)

The speed of foraging is denoted by $Z_{g}$,$\psi_{g}$ is the inertia weight and $G_{k}^{Old}$ defines the last movement of foraging. The effect of best fitness is defined by $\zeta_{k}$ and is determined using Eq. (20).

$$\zeta_{k} = \zeta_{k}^{Food} + \zeta_{k}^{Bst}$$

(20)

The term $\zeta_{k}^{Food}$ is the food attractive and $\zeta_{k}^{Bst}$ is the best fitness function. The center of food is evaluated using Eq. (21).

$$D^{Food} = \frac{{\sum\limits_{k = 1}^{S} {\frac{1}{{I_{k} }}D_{k} } }}{{\sum\limits_{k = 1}^{S} {\frac{1}{{I_{k} }}} }}$$

(21)

The food attraction $\zeta_{k}^{Food}$ is examined in Eq. (22).

$$\zeta_{k}^{Food} = Q^{Food} \hat{I}_{k,Bst} \hat{D}_{k,Food}$$

(22)

The term $Q^{Food}$ is the food coefficient, and it is evaluated using Eq. (23).

$$Q^{Food} = 2\left( {1 - \frac{It}{{M_{It} }}} \right)$$

(23)

The best fitness is described in Eq. (24).

$$\zeta_{k}^{Bst} = \hat{I}_{k,Bst} \hat{D}_{k,Bst}$$

(24)

Here, the term $\hat{I}_{k,Bst}$ denotes the best position to be previously visited.

Physical diffusion

It is a random process given in Eq. (25).

$$F_{k} = F^{Mx} \kappa$$

(25)

It depends on the factors of the random directional vector $\kappa$ and the maximum diffusion speed $F^{Mx}$. If the iteration number is increased, it linearly decreases, and it is formulated in Eq. (26).

$$F_{k} = F^{Mx} \left( {1 - \frac{It}{{M_{It} }}} \right)\kappa$$

(26)

Movement operation

The movement of the operation mainly depends on the scaling factor $\Delta m$. It completely depends on the candidate space, and it is represented in Eq. (27).

$$\Delta m = Q_{m} \sum\limits_{q = 1}^{SZ} {\left( {upr_{q} - lwr_{q} } \right)}$$

(27)

The total number of variables is denoted by $SZ$, “the upper and the lower bounds are represented by $upr_{q}$, and $lwr_{q}$, respectively.” The pseudocode of the developed HZKHO is described in Algorithm 1. The flowchart of the HZKHO is indicated in Fig. 5.

Algorithm 1

Offered HZKHO.

Fig. 5

Flowchart of the developed HZKHO.

Weighted fused feature selection

The extracted deep features $Df_{z}^{VGG}$, color features $CL_{a}^{fs}$, shape features $SP_{b}^{fl}$_, and texture features $TX_{c}^{fx}$ are fed to the weighted fused feature selection stage. Fifty features, including color, shape, and texture, is optimally selected using the developed HZKHO. Then, an optimization algorithm is applied to optimize the weights, which are multiplied by the selected features to obtain the fused features. The major objective of using this weighted fused feature selection is to enhance the overall performance of the developed model. By using this method, the least significant features are eliminated. Therefore, only the most relevant features are retained, ensuring no significant information is lost. Moreover, the dimensionality is effectively reduced through this fused feature selection process. It also helps to eliminate information redundancy and feature overlap. The structural correlation among features is captured by using this weighted fused feature selection. The selected features after applying the weighted fusion strategy are denoted by the term $CS_{b * }^{Wf}$. The schematic representation of the proposed weighted fused feature selection using HZKHO is shown in Fig. 6.

Fig. 6

Schematic representation of the developed HZKHO-based weighted fused feature selection process.

Medicinal Plant leaf disease classification using attention-based dilated adaptive densenet over the selected weighted fused features

Dilated adaptive densenet

For recognizing large-scale image fields with high efficiency, we use densely connected networks. It effectively reduces the gradient vanishing problems. The layers are directed and integrated into the DenseNet, which guarantees the transmission of information among all the layers. The DenseNet⁴⁶ is composed of several basic units, where the $p^{th}$ input unit is represented as $\left( {CS^{\left( p \right)} } \right)$_, and it contains the outputs of the preceding block $\left( {1,2,...,p - 1} \right)$. It is the first basic unit that is given to the input of the next connected basic unit. By considering the appropriate range for the non-linear activation function, the input data is scaled in the batch normalization layer. With the usage of the ReLu non-linear activation function, the expression ability of the DenseNet is improved. The batch normalization equation is expressed in Eq. (28).

$$I^{\left( p \right)} = \kappa \times \frac{{CS^{\left( p \right)} - Mean\left[ {CS^{\left( p \right)} } \right]}}{{\sqrt {VAR\left[ {CS^{\left( p \right)} } \right]} }} + \vartheta$$

(28)

The terms $\kappa$ and $\vartheta$ are the scaling factors and the shifting factors. Moreover, the batch normalization factor effectively eliminates the covariance shift that occurs internally, and it maintains a stable data distribution state.

In the DADensenet, the parameters such as the activation function, hidden neuron count, and epochs are optimized within a particular range. This optimization improves the classification efficiency while considering precision, false positive rate, and negative predictive value. The optimization process is carried out with the use of the developed HZKHO. The systematic representation of Dilated Adaptive DenseNet is shown in Fig. 7.

Fig. 7

Schematic illustration of the dilated adaptive DenseNet model.

Proposed attention-based dilated adaptive DenseNet

The proposed A-DADensenet is used in the “classification of a medicinal plant leaf disease model” to classify the leaf images very effectively.

Initially, the weighted fused selected features $CS_{b * }^{Wf}$ are inputted into the DADensenet to effectively learn the discriminative features from the images. Here, features are organized into dense blocks, where layers are densely connected. This allows for efficient feature extraction and reduces the vanishing gradient problem. The dilated convolution expanded the kernel size by inserting a hole between the consecutive elements in the dilated convolution process. The usage of receptive fields in the dilation block is usually larger than the naïve convolution, and hence, it learns more features from the images. Here, the skipped connections are introduced to find the features from the unrecognized areas. The usage of the attention mechanism is it gives special attention to particular regions during the recognition process. An attention mechanism is used to adjust the particular weights for the features. The self-attention network is expressed in Eq. (29).

$$Soft_{Max} \left( {\varpi^{T} * y + \gamma } \right)$$

(29)

The softmax activation function is indicated by $Soft_{Max} \left( \cdot \right)$ in the attention network, the weights are indicated by $\varpi$ and the bias is denoted by $\gamma$. The transpose operation is expressed by $T$. The input vector with $M$ elements is indicated by $\delta = \left( {\varpi^{T} * y + \gamma } \right)$ and the $n^{th}$ attention output is indicated in Eq. (30).

$$\lambda_{n} = Soft_{Max} \left( {\delta_{n} } \right) = \frac{{\exp \left( {\delta_{n} } \right)}}{{\sum\limits_{n = 1}^{M} {\exp \left( {\delta_{n} } \right)} }}$$

(30)

The final attention weights or the outputs are indicated by $\lambda_{n} = \left[ {\lambda_{1} ,\lambda_{2} ...,\lambda_{M} } \right]$. Then, these weights are multiplied by the features with the help of element-wise multiplication that as given in Eq. (31).

$$\vec{F} = f \circ \lambda$$

(31)

In addition, the parameters from the DenseNet are optimized, including the activation function and hidden neuron count, and the epochs are optimized with the utilization of recommended HZKHO. This may improve the classification performance over medicinal plant leaf diseases. “The objective function of the framework is explained in Eq. (32).”

$$OT = \mathop {\arg \min }\limits_{{\left\{ {He_{x}^{DN} ,AT_{y}^{DN} ,Eo_{z}^{DN} ,Wt_{v}^{Fs} ,FS_{r}^{fs} } \right\}}} \left( {\frac{1}{prcn} + \frac{1}{Npv} + Fpr} \right)$$

(32)

Here, the term $He_{x}^{DN}$ is the optimized hidden neuron count in the interval of $\left[ {5,255} \right]$, $AT_{y}^{DN}$ is the optimized activation function in the interval of $\left[ {1 - 4} \right]$, $Eo_{z}^{DN}$ is the optimized epochs from DenseNet in the range of $\left[ {5,50} \right]$_, and $Wt_{v}^{Fs}$ is the optimized weights in the range of $\left[ {0.01,0.99} \right]$_, and $FS_{r}^{fs}$ is the optimally chosen features in the range of $\left[ {1,50} \right]$. Moreover, the term $prcn$ is the precision measure,$Npv$ is the negative predictive value, and $Fpr$ is the false positive rate. The formula for estimating the effectiveness measure is given in Eqs. (33), (34) and (35), respectively.

$$prcn = \frac{{Neg_{Tr} }}{{Pos_{Tr} + Pos_{Fl} }}$$

(33)

$$Fpr = \frac{{Neg_{Fl} }}{{Pos_{Fl} + Pos_{tr} }}$$

(34)

$$Fpr = \frac{{Neg_{Tr} }}{{Neg_{Tr} + Neg_{Fl} }}$$

(35)

“Here, the term $Neg_{Tr}$ is the true negative,$Neg_{Fl}$ is the false negative,$Pos_{Fl}$ is the false positive, and $Pos_{Tr}$ is the true positive.”Finally, the suggested A-DADensenet offered the classified medical plant leaf disease outcome. The developed A-DADensenet-based disease classification framework from medicinal plant leaves is given in Fig. 8.

Fig. 8

Developed A-DADensenet-based medicinal plant leaf disease classification model.

A newly developed deep learning-based disease classification method can effectively classify the types of medicinal plant leaf diseases at a very early stage. Compared to existing models with proven higher accuracy, the proposed A-DADenseNet model and HZKHO optimization algorithm offer several advantages. For instance, existing hybrid deep learning methods²³ achieved 98.69% accuracy. “However, it may rely on manual feature extraction, which can be time-consuming and may not capture all relevant features.” In contrast, the A-DADenseNet model’s attention mechanism enables automatic feature extraction, reducing the need for manual intervention. In addition, existing deep CNNs attained 99.64% accuracy⁴⁷ in the classification process, yet they can get stuck in local optimal solutions that sometimes affect the model’s performance. In contrast, HZKHO is an optimization algorithm that focuses on minimizing the loss function within a required time while avoiding local extremes and achieving rapid convergence in the global minimum. This helps A-DADensenet learn more efficiently and find optimal parameters for higher accuracy, especially when dealing with complex datasets. Even though the existing machine learning model⁴⁸ attains a high accuracy of 99.01%, they have difficulty classifying the leaf disease from different variations like color, shape, and texture. But the developed model used the Multi-Scale VGG16 to extract the deep features, and the other features, such as color, shape, and texture, are separately retrieved from the segmented images. This process makes the classification process easier. The combination of automatic feature extraction and adaptive hyperparameter optimization enables the A-DADensenet model to generalize better to new, unseen data. This is particularly important in real-world applications, where the model may encounter data that differs significantly from the training data. Overall, the proposed A-DADensenet model and HZKHO optimization algorithm provide a powerful tool for medical leaf disease classification, with significant potential for real-world applications.

Validation metrics

Several positive and negative validation metrics are utilized in the proposed method to analyze the performance, as described below.

$$MCC = \frac{{Pos_{{Tr}} + Pos_{{Fl}} + Neg_{{Fl}} - Pos_{{Fl}} }}{{\sqrt {\begin{array}{*{20}c} {\left( {Pos_{{Tr}} + Pos_{{Fl}} } \right)} & {\left( {Neg_{{Tr}} + Pos_{{Fl}} } \right)} \\ {\left( {Pos_{{Tr}} + Neg_{{Fl}} } \right)} & {\left( {Neg_{{Tr}} + Neg_{{Fl}} } \right)} \\ \end{array} } }}$$

(36)

$$f1 - measure = \frac{{2Pos_{Tr} }}{{2Pos_{Tr} + Pos_{Fl} + Neg_{Fl} }}$$

(37)

$$Sensitivity = \frac{{Pos_{Tr} }}{{Pos_{Tr} + Neg_{Fl} }}$$

(38)

$$FNR = \frac{{Neg_{Fl} }}{{Pos_{Tr} + Neg_{Fl} }}$$

(39)

$$Specificity = \frac{{Neg_{Tr} }}{{Neg_{Tr} + Pos_{Fl} }}$$

(40)

$$FDR = \frac{{Pos_{Tr} }}{{Neg_{Tr} + Neg_{Fl} }}$$

(41)

$$Accuracy = \frac{{Pos_{Tr} + Neg_{Tr} }}{{Pos_{Tr} + Neg_{Tr} + Pos_{Fl} + Neg_{Fl} }}$$

(42)

Results and discussions

Experimental setup

The advanced deep learning-based medicinal plant classification approach has been designed in the software platform Python. Here, the experimental analysis has been taken to ensure the classification performance of the designed model concerning different positive and negative performance measures. The population was assumed as 10, the chromosomal length was 3, and the number of iterations was 25 in the designed model used to analyze the performance. Here, several heuristic algorithms were considered for performing the comparative analysis, and they were the Deer Hunting Optimization Algorithm (DHOA)⁴⁹, the African Vulture Optimization Algorithm (AVOA)⁵⁰, the ZOA⁴⁴, and KHO⁴⁵. The classification approaches that were recently used for the classification of medicinal plant leaf diseases were ResNet⁵¹, VGG16⁵², DenseNet⁵³, and A-DADenseNet⁴⁶.

Resultant preprocessed segmented medicinal plant leaf images

The resultant preprocessed images and the segmented leaf images are given in Fig. 9.

Fig. 9

Resultant preprocessed and segmented images.

Convergence and ROC analysis of the presented framework

To evaluate the performance of the designed HZKHO-A-DADensenet-based medicinal plant leaf classification approach, the cost function and the ROC metric are considered. Figure 10, given below, displays the convergence analysis of the designed approach in terms of the cost function among various algorithms. The convergence graph shows that the cost function value is slightly decreased when increasing the number of iterations. The cost function is very low in the developed model when compared with other algorithms. The AUC analysis shows that the designed method gained with enhanced accuracy of 11.68%, 9.01%, 5.87%, and 2.78% superior to the techniques like ResNet, VGG16, DenseNet, and A-DADensenet, for the false positive rate at 0.3. The accuracy of classification is highly improved in the recommended model than the other algorithms.

Fig. 10

Classification analysis of the proposed medicinal plant leaf disease classification framework (a) Cost function and (b) ROC.

Extracted features based performance analysis of the proposed method

The extracted features based performance examination of the designed method over different optimization algorithms is represented in Fig. 11, and various recent methods are indicated in Fig. 12. This analysis is used for all the extracted features among various datasets. From the graph, the developed HZKHO-A-DADensenet-based disease classification approach from plant leaves attained with greater accuracy of 10.02% than DHOA-A-DADensenet, 1.93% than KHO-A-DADensenet, 17.81% than AVOA-A-DADensenet, and 15.12% than ZOA-A-DADensenet for considering the texture dataset. The classification accuracy for all datasets is higher in the developed model than in other algorithms and techniques.

Fig. 11

Extracted features-based performance analysis on a developed disease classification model from medicinal plant leaves among various algorithms for the measures as (a) Accuracy (b) FDR (c) F1score (d) NPV (e) FNR (f) Precision (g) FPR (h) Sensitivity (i) MCC and (j) Specificity.

Fig. 12

Extracted features-based performance analysis on developed advanced deep learning-based disease classification model from medicinal plant leaves among various techniques for the measures as (a) Accuracy (b) FDR (c) F1score (d) NPV (e) FNR (f) Precision (g) FPR (h) Sensitivity (i) MCC and (j) Specificity.

Overall comparison analysis of the developed method

The analysis of the suggested model over various algorithms is represented in Table 2, and previous medicinal plant leaf classification schemes are illustrated in Table 3. From the tabular analysis, the recommended model achieved with greater NPV of 6.14% than DHOA-A-DA-Densenet, 4.96% than KHO-A-DADensenet, 6.12% than AVOA-A-DADensenet, and 8.36% than ZOA-A-DA-Densenet.

Table 2 Performance comparison of the designed method among various algorithms.

Table 3 Performance comparison of the developed model among various methods.

Statistical analysis of the introduced method

The statistical evaluation of the proposed model is given in Table 4 below. The statistical evaluation enhances the presented method’s performance by using the statistical metrics. From this evaluation, the developed model attained a highly improved mean, median, and standard deviation value than the others.

Table 4 Statistical examination of the developed model among various algorithms.

Overall efficacy validation of the designed method using recent models

The overall efficacy of the presented medicinal plant leaf classification schemes using recent approaches and state-of-the-art models is given in Tables 5 and 6. The accuracy and NPV rate of the designed model were 89.14 and 89.21%. Thus, the result shows that the developed method provides superior outcomes over recent baseline frameworks.

Table 5 Overall efficacy of the developed model using recent approaches.

Table 6 Performance analysis of the designed method over state-of-the-art models.

Advantages and disadvantages of the designed medicinal plant leaf disease classification

The advanced features of the designed model are discussed as follows. A new medicinal plant disease classification method is implemented for recognizing the diseases. This performance improvement of the designed model applies to clinical and medical applications. The HZKHO algorithm and A-DADensenet model are introduced for optimization and classification purposes. The optimization of parameters in the DenseNet model is utilized for resolving the overfitting issues. The simulation findings of the designed model have attained a better convergence rate over other approaches it can be utilized to exploit gradient issues. The statistical analysis of the designed model revealed that it is statistically significant. However, the designed model needs more evaluation studies to discover the extent of the algorithm’s capability to solve optimization problems. The major weakness of the designed method is in validating the comparative studies with complicated test functions. In the future, the designed model will add more evaluation studies, which will be utilized to resolve the discrete optimization problems and various complex problems.

Conclusions

A novel disease classification model for medicinal plant leaves has been developed using an advanced deep learning algorithm to accurately identify the types of diseases affecting medicinal plants. The collected leaf samples were first preprocessed, and the resulting high-quality images were then subjected to adaptive-thresholding-based segmentation. Next, the segmented images were passed through the feature extraction stage, where relevant features were extracted. From these, weighted fused features were selected using the proposed HZKHO algorithm. Finally, disease classification was performed using the Attention-based Dilated Adaptive DenseNet, with its parameters optimized through the same HZKHO approach. The results were compared with existing models to assess effectiveness, and the outcomes demonstrated that the specificity of the proposed HZKHO-A-DenseNet-based medicinal plant leaf disease classification model was improved by 4.02%, 0.45%, 0.96%, and 0.51% over ResNet, VGG16, DenseNet, and A-DADensenet, respectively. It delivers superior classification performance compared to existing algorithms and previous studies.

Limitations and future scope

The developed model has difficulty in handling large amounts of high-quality images. Also, it has limited adaptability to different conditions. Therefore, future work will use the data augmentation technique to improve models’ ability to recognize diseases under different visual conditions. Moreover, to enhance the image quality, the data annotation method will be used for further research. This technique will allow the developed model to identify and classify images accurately, leading to reliable outcomes.