Adaptive hybrid segmentation combined with meta heuristic optimization in transfer learning for plant leaf disease classification

Abstract

Plant diseases can damage specific parts of leaves for better readability during the farming process. Plants refer to various types of crops, including fruits and vegetables. During the production phase of healthy crops, plant diseases often begin by infecting the leaves. Leaves, being exposed, are more vulnerable to disease than other plant parts. When affected by disease, crop yield decreases, leading to economic loss. Hence, early disease identification model is required and deployed in an automated computerized way. The analysis have shown that multiple approaches were executed to detect the disease, still it suffers from pitfalls like inadequate feature extraction, handcrafted features, computation burden, complexities and so on. To improve the process, an efficient method is developed for detecting various plant diseases by different learning method. Firstly, the different plant leaf data were gathered from UCI, Kaggle web sources and benchmarks. The unwanted noise in the input leaf images are pre-processed by using median filter. Subsequently, the affected or abnormal region was segmented by the adaptive hybrid K-means with fuzzy C-means clustering (AHKM-FCM); the parameter tuning is also done by improved random variable-based water strider algorithm (IRV-WSA). Finally, the segmented region was subjected into the Transfer Learning Network that was processed with Efficient-net, ResNet and Densenet, in which fine tuning of weight was accomplished by using the IRV-WSA. The model was analyzed and computed across divergent measurements. Classification output results from the proposed IRV-WSA-ETLNet model include 94.853% accuracy, 94.750% sensitivity, 94.888% specificity, and 96.068% F1 Score. Additionally, this system uses less computing time 24.378 ms. Compared to previous methods, the findings of the proposed model demonstrate improved classification rates and help the farmer to increase crop production.

Introduction

Farming is the primary need of the country. Flora is the more important source for the survival of humans and it is critical to care for them¹. For successive agricultural development, there is a need to investigate the unhealthy and healthy plant, which is more significant².

Many plant leaf diseases can totally affects the production of the crop field, thus it creates a huge loss in economic to all farmers. In general, the diseases in a plant is naked with various learn and skill to observe the manifestations of diseases in plants³. More importantly, the nuts, vegetables, legumes and grains productivity and qualities are infected by various factors in plants⁴. Hence, the leaves of the plant are the important source that affect with certain disease caused by insects or other factors. Most of the diseases in plants show visible manifestation, and also the method accepted recently is the experienced plant pathologist treatment of the diseases via optical inspection of the infected leaves of the plant⁵. Disease classification is more suggested process for identifying the plant infections by classifying various manifestations of such infections⁶. Plant diseases, which frequently start as infections on the leaves, have a major effect on crop health and agricultural output. These illnesses can spread quickly if they are not identified and treated in a timely manner, which might result in lower agricultural yields and financial losses. Conventional techniques for identifying diseases depend on manual examination, which is labor-intensive, time-consuming, and prone to errors. Currently used automated illness diagnosis methods have several drawbacks, including poor feature extraction, a dependence on manually created features, expensive calculation, and implementation complexity.

A plant disease detection system that is automated, effective, and precise is desperately needed. It must be able to examine leaf photos, diagnose illnesses early, and let farmers take prompt preventive measures.

The importance of an automated plant disease identification systems are,

• Farmers can achieve higher yields and quality, minimizes economic losses, and reduces dependency on manual inspection.

• Advanced methods like machine learning and deep learning should be used by the system to improve detection accuracy while reducing operational complexity and computing load.

• Accurate and timely diagnosis the diseases in plants is the main significance for managing the appropriate and sustainable agricultural fields, and also preventing the unwanted wastes of other resource and finances.

Several diseases in plants don’t have visible manifestations, and it’s unavoidable to utilize highly developed analysis techniques in some diseases⁷. Several limitations still remain in current methods, despite progress in automated plant disease diagnosis. These include low feature extraction, low disease classification accuracy, high computing burden, and reliance on good-quality data, restricted real-time monitoring, and automation. The process of extracting the features while classifying the different plant leaf diseases is a quite challenging task since it emerges with beneficial key points⁸. In the conventional way of deploying such networks or approaches, the feature vector representation occupies the pivotal space. Still, the existing methodologies are facing the issues of utilizing handcrafted features⁹. This kind of feature leads to cause the performance degradation in learning methodologies. Moreover, the categorization of various leaf types with respect to different plants is becoming cumbersome. Also, the large-scale dimensional data has been managed difficulty while identifying the disease¹⁰. In general, the training and testing phase are the two major functions in network. Since multiple data is used for phases, the computation burden or time complexity might occur¹¹. Specifically, certain proficient methods with a few parameters trained quickly and no need for compromise on performances is unavoidable¹².

A more effective and precise model is created with cutting-edge learning techniques in order to get beyond the drawbacks of the current plant disease detection methods. In order to increase the accuracy of illness diagnosis, the suggested method combines segmentation, feature extraction, and deep learning-based classification with optimal parameter adjustment. In recent times, the deep learning (DL) approach is familiar in detecting crop diseases. Artificial Intelligence (AI) techniques are frequently used as a treatment for plant diseases with computer knowledge¹³. The Artificial intelligence techniques are the group of methods with the help of deep learning approach, this process is carried and there is no need for characteristic arrangements¹⁴. Deep learning-based methods are utilized for plant disease identification on leaves. This investigates the leaf images taken from various places and categorizes the leaves in one of the set of predefined classes¹⁵. The implementation of a perfect approach based on deep learning may utilize to recognize the leaf disease. Datasets are trained based on deep learning approaches to provide accurate results¹⁶. Next, testing phase are executed to evaluate the exhibition of the task by some performance measures. The deep learning research appears to have many possibilities that enhance the accuracy. The considerable expansions and advancements in the deep learning method have been developed to improve the accuracy to identify disease in plant leaves¹⁷.

The significant involvement of the plant leaf disease identification model as given as follows:

• To enhance the plant leaf disease identification model by utilizing AI techniques to support the agricultural fields and farmers to provide healthy food to all individuals.

• To implement the segmentation process, the AHKM-FCM model is developed to get an accurate segmented region of affected parts, and its parameters are enhanced by IRV-WSA approach to increase the accuracy.

• To execute the classification process, the Efficient Transfer Learning Network (ETLNet) model is developed by assembling the classifiers such as EfficientNet, ResNet and DenseNet.

• To improve the mechanism, a new algorithm named IRV-WSA is developed for tuning the parameters and to provide more accurate classification results. In the segmentation process, the tuned parameters are the number of attempts in KMC and FCM to improve the variance and correlation coefficient. In the classification process, the tuned parameters are weights in Efficientnet, Resnet and Densenet for increasing of accuracy, NPV, MCC and also reducing the FNR to give classified outcome.

• To determine the segmentation and classification performance of the recommended model, by comparing with other conventional models.

The various section of proposed system is stated as follows. The 2nd part shows the reference concepts for the proposed model. The 3rd section explains the development of adaptive hybrid segmentation with transfer learning approach for the model. The 4th section describes the AHKM-FCM clustering based plant leaf disease segmentation. The 5th section illustrates the result and discussion of the developed model and 6th section provides the summary of the model.

Literature survey

Related works

A model developed for the leaf disease detection and its classifications¹⁸. Initially, the model recognized the leaf disease and its type using DenseNet framework. The Densenet technique trained the images and categorized them based on their nature. This method was utilized for checking the newly taken images of a leaf. The developed DenseNet architecture generated higher accuracy in detecting, and a minimum number of leaf images were utilized in the training period. Next, in stage two, semantic segmentation method based on deep learning has perceived the damages in leaves. Every value of pixel in the leaf image was separated and training supervision was acted on the pixel value using one dimensional convolutional neural network (CNN). Then the model was proficient to identify the disease in the leaves at each level. The valuation of the developed semantic segmentation model provided 97% of accuracy. Stage three recommended the solution for leaf disease according to the damaged state and type of the disease. The developed model classified different infections in various plants during the analysis, such as apple, potato, strawberries and grapes. The developed method was analyzed with conventional methods and provided the best performances.

The classification of diseases in Oryza sativa by using the various Machine Learning (ML) technique, the recommended model has reduced the dependency of existing techniques to prevent the paddy crop from the diseases namely, false smuts, bacterial leaves blight, brown leaves spot, sheath rots and rice blasts, which were the 5 basic diseases, have often damaged the rice in Indian field. The pre-processed image and segmented image have utilized to identify the section of diseased in paddy plants, with the above listed diseases being classified correctly based on the visual content of them. The integration of Support Vector Machine (SVM) classifiers and CNN were used to group and recognize the particular type of disease in paddy plants. The recommended deep learning-based task with softmax function and ReLU achieved the maximum evaluation accuracy. The predictive solution suggested in this paper provided the agriculture-correlated organizations and individuals in taking the appropriate metrics to struggle against paddy diseases¹⁹.

An optimized approach to identify and classify the leaf disease using Ant colony Optimization technique with CNN was designed²⁰. The leaf surface, color features and its arrangements were obtained from the original image by CNN. Few effective measures were utilized for investigation and developing the recommended technique that confirmed the developed task worked more than other existing methods for proving more accuracy. The stages were utilized in process of detecting the disease was, separating the image, removing noise, classification and pictures acquisition.

The EfficientNet model to group the diseases in plant leaves, which are compared with different deep learning methods. The leaves of village plants were utilized for the trained model. Augmented and original datasets trained the method had 61,486 and 55,448 leaf images. The other deep learning and EfficientNet techniques have been training by utilizing the transfer learning model. Results provided by the EfficientNet models attained the maximum accuracy value in augmented and original datasets of 99.97 and 99.91% and achieved a precision of 99.39 and 98.42%, respectively²¹.

The deep learning method developed to classify leaf disease²². Initially, a step in pre-processing was acted using the method named as Median Filtering. This method avoided the noise present in leaf images. After the step of image pre-processing, they introduced the Binary patterns to extract the features in leaf. With CNN model, they grouped and detected the diseases easily. The decision support system (DSS) helped the farmers to employ the program for effectual treatments. These accepted the farmers for increasing the proficiency of the control methods without raising the challenges. With the help of MATLAB Simulink application, this technique had been executed and evaluated. The developed method performed best compared to other conventional techniques like Deep CNN, and VGG-INCEP.

A proficient method for detecting disease in rice plants regarding CNN task. The three familiar diseases in rice were brown spots and leaf smuts and the bacterial leaf blights, focused in this method. The model based on the size, color and shape of the leaves images. The developed method used Otsu’s global threshold techniques to act as binarization of images to separate the background noises present in leaf images. The developed model regarding the fully connected CNN model used 400 sample images of leaves with diseases during trained period and also 400 samples images of healthiest leaves for detecting the above mentioned three diseases in Oryza sativa. The experiment proved that the proposed CNN method was fast and efficient, which provided 99.7% accuracy. The same was compared with existing plant disease classification and recognition task²³.

The proficient deep CNN technique since it was identified the significant factors automatically without the need for the supervision of the human. With the support of deep learning based models, DCNN deeply classified the diseases in plants. Additionally, the training and testing process of DCNN proved the correct classification and determination result of the diseases present in groundnut. The sample images of groundnut leaves with disease were taken from village plant datasets, and it utilized for a process of testing and training. They used different models to train datasets, and it showed the best performances of the developed DCNN model. From overall analysis, it provided the accuracy value of 99.88% in developed DCNN model²⁴.

Based on machine learning techniques for detecting the leaf disease in grape leaves and correctly differentiated with disease classes²⁵. Moreover, the developed model is to classify the grape leaf disease. The histogram with more quality and modified features in histogram were utilized for proving the boundary, discriminative data, patterns and structures. Then, the process of classification was acted based on obtaining more quality gradient features. The accuracy of the classification process was enhanced to a huge extent by utilizing the developed methods. The developed methods showed the best accuracy value compared with conventional techniques.

The efficacy of identifying and classifying rice plant diseases is the primary focus of this study. The AX-RetinaNet model’s hyperparameters are tuned using the CAHA optimization model. In order to categorize rice plants as either healthy or sick, this study incorporates three different types of disease detection datasets: rice plant, rice leaf, and rice disease. To verify the efficacy of illness detection, the most crucial performance indicators are precision, F1-score, accuracy, specificity, and recall. Comparing the suggested CAHA-AXRNet methodology to other current rice plant disease detection techniques, it shows efficacy and achieves a higher accuracy rate²⁶.

A DCNN (Dilated convolutional neural network) model with global average pooling (GAP) was created²⁷. It will be built by substituting dilated convolution kernels for standard CNN convolution kernels and by using global average pooling in lieu of the fully connected layer in conventional CNN. Reduced memory use and computational cost are benefits of the dilated convolution, while overfitting is prevented by GAP. Using performance metrics including accuracy, precision, recall, and f1-score, the outcomes of this method are compared with different learning and hybrid learning approaches. These two novel ideas are implemented using CNN. Performance measurements indicate that, when compared to the standard CNN model, the DCNN model with GAP improves training accuracy by an average of 5.5%. The outcomes are compared with those of other learning and hybrid learning approaches.

This work introduces a sophisticated LSTM-RNN model designed to forecast potato prices. Five conventional machine learning models were employed to compare the model’s performance: K-nearest neighbor, random forest, support vector regressor, linear regression, and gradient boosting regressor were used to categorize isolated households and ascertain their socioeconomic standing. The actual evidence suggested that our suggested LSTM-RNN model outperformed all comparative models in terms of efficiency. In order to plan and implement solutions to address food security, this model incorporates price stability, which is essential to a balanced economy and a sustainable food system²⁸.

The authors discuss prevalent illnesses and the several stages of research on such detection systems. They analyze existing feature extraction techniques to identify the one that appears to work best across a range of crop types. Researchers’ comprehension of the application of computer vision to the detection and classification of plant diseases would be improved by this work²⁹.

Problem statement

The condition of a plant’s systemic abnormal physiological function, those results from the ongoing, protracted “irritation” produced by phyto-pathogenic organisms are referred to as plant disease. The main substance of the leaf disease can be generally categorized as noninfectious and infectious. Infectious plant illnesses are caused by infectious agent. To categories the various leaf diseases, a DL model is built. Some of the research challenges are noted as follows:

• The leaf images are the major source of finding the plant disease, yet the background images are indulged in reducing the performance efficiency. While processing with unwanted background factors, the system faces such complexities that to be rectified.

• Through the traditional methods, the desired results have obtained. However, the diversity of large datasets or used data source is limited. The detection also becomes critical in state of complex environment.

• Some of the methods processing is still cumbersome due to time consumption, high subjective, and more labor consuming. Being relied on artificial feature space representation, the system can mislead the leaf disease detection process.

• While capturing the data, its characteristics may get changed or altered. Like, less lighting illumination, more blurry regions are the challenging factor in forecasting the plant leaf diseases.

• There is a view of deploying minimal number of automated models to classify the plant diseases. Hence, designing a novel method is suggested with the application of intelligent or learning models.

A significant impact on crop health and agricultural productivity is caused by plant diseases, which often begin as infections on the leaves. Lower agricultural output and financial losses may arise from these diseases’ rapid spread if they are not detected and treated in a timely way. The manual inspection used in conventional illness identification methods is time-consuming, labor-intensive, and error-prone. Poor feature extraction, a reliance on manually generated features, costly computation, and implementation complexity are some of the shortcomings of the automated sickness detection techniques now in use.

Development of adaptive hybrid segmentation and transfer learning

For solving all existing issues, we developed a new model based on deep learning. Figure 1 depicted the general structure of proposed methodology.

Fig. 1

Structure of proposed methodology.

Data collection

The image dataset is accessible on the plant village website, notably for images of plant leaf disease. The training and testing of the disease classification is done using these labelled images. The developed plant leaf disease classification model has gathered the diseased and healthy leaf images from the standard dataset.

Dataset 1 (“PlantifyDr Dataset”): The essential images are gathered from the PlantifyDr dataset at the link https://www.kaggle.com/datasets/lavaman151/plantifydr-dataset: access date: 2024–07-26. This dataset has nearly 125,000 images of 10 various types of plants and they are, bell, apple, cherry, pepper, citrus, corn, peach, grape, potato, peach, tomato and strawberry.

Pre-processing

Pre-processing is the technique used to raise the calibre of the input image. It is an important step in the recognition process since a noisy image has a direct impact on how diseases are categorized. This method uses three steps to enhance an input image of plant leaves: denoising, contrast correction, and edge enhancement. After pre-processing, effective and proficient picture division is essential for producing precise differentiating proof using the input images. Here, the median filter is used for pre-processing. Pre-processing is done on a provided image to improve quality and eliminate unwanted distortion in the input leaf image. To indicate the value of each pixel, an RGB (Red, Green, and Blue) matrix is utilized. The purpose of the pre-processing stage is to enhance the clarity and contrast of the image by removing noise and variance from the plant’s leaf image. In addition to contrast enhancement and noise reduction, the pre-processing stage can be utilized for image normalisation and non-uniform intensity correction to eliminate artefacts and enhance accuracy of the subsequent processing steps. The image characteristics are also localised, retrieved, and segmented from pictures for additional classification in pre-trained models.

Image segmentation

After the pre-processing, the leaf image is given into the segmentation stage. Each image is divided into a huge number of pieces or regions during segmentation. The application and user requirements have the biggest impact on the number of pieces. AHKM-FCM is used for segmenting the plant leaves. The image is sharply and precisely segmented during image segmentation. The result of image segmentation is a piece of a picture, which is made up of a deleted set of the image’s contours. The main goal of segmentation is to effectively change or simplify the representation already existent in the image for carrying out the evaluation. Image segmentation divides the image effectively for additional processing. The process of photo segmentation assigns a value to each pixel. In this segmentation, the leaf portions are segmented as affected and unaffected.

Feature extraction

Once the segmentation was done, the segmented image is fed into the feature extraction process. The primary objective of the feature extraction is to transform the image data into a format that makes it easier to match leaf images. The two steps that make up this feature extraction and feature selection. The feature extraction process looks for several traits that best describe an image of plant leaves. Because there are generally many features chosen, a feature selection algorithm is used in the second stage to choose the most noticeable features. After segmentation, its takes place for feature extraction and it is done by using 13 features of leaf images. Five categories of features were extracted during feature extraction. They are leaf -related features, geometric features, colour features, texture features, and fractal features. The majority of plant leaf images identification research solely considers leaf, colour, and texture aspects. In this investigation, the geometric and fractal characteristics were also considered. These characteristics were retrieved and divided into three groups, including shape, colour, and texture. The classifier receives these features as input for additional processing. The technique is simply carried out to provide a clear comparison between the extracted results and the data set, for the anticipated occurrence of impacted diseases in the leaf.

Classification of leaf image

The final stage of the work is classifying the plant to which the input leaf image belongs. The algorithm’s core consists of an iterative recognition process that contrasts properties from the input image with feature vectors corresponding to images of plant leaves from the pre-built dataset. This is essential since even one diseased plant can transmit the disease. Furthermore, most farmers are not sufficiently informed about these ailments and efficient remedies. The cost of employing specialists could be exorbitant, and using pesticides improperly will harm the ecosystem. Because of this, we have created a newly designed solution to address this issue.

The system taken different plant leaf images datasets as input. The final stage is leaf disease classification, in this process the integrated technique is used in our suggested system. The parameter tuning is also done by Improved Random Variable-based Water Strider Algorithm (IRV-WSA). Finally, the region was subjected into the Transfer Learning Network that was processed with Efficient-net, ResNet and Densenet, in which fine tuning of weight was accomplished by using the IRV-WSA.

Performance evaluation

The confusion metrics used to evaluate the overall performances by using the various measuring parameters. The model was analyzed and computed across divergent measurements. The subsequent measures investigate the operation of the proposed model. Compared to existing methods, the outcomes of the proposed IRV-WSA-ETLNet model demonstrate improved classification rates.

Implemented plant leaf disease classification framework

The segmentation process helps to get an actual diseased leaf and carried out for disease classification. The classification process is used to identify the diseased leaves quickly to support individual person. So, here we developed the model of leaf disease segmentation and its grouping. The pictorial illustration of the developed model is described in (Fig. 2).

Fig. 2

Pictorial illustration of the proposed model.

The developed model is based on the deep learning approach which is used to detect the leaf disease early and support the farmers and the people. Initially the leaf images are collected from the farming land using the 50MP Primary Camera with 2MPMacro Lens. Then resize the image into 256 × 256 pixels. The unwanted distortions in the given image are filtered (pre-processed) by using the median filter. Further the input data is carried to the process of segmentation process. In this, the AHKM-FCM model is used to divide the diseased region from the input data. Here, the parameters in the developed AHKM-FCM technique are tuned by IRV-WSA algorithm to increase the segmentation performance.

The parameters tuned in this segmentation process such as number of attempts in KMC, fuzziness parameters in FCM, iteration in FCM and epsilon in FCM to increase the variance and correlation coefficient. After completing the segmentation process, it provided segmented images as the output. Then, these segmented images are taken to the classification operation with the help of implemented ETLNet model. The ETLNet model comprises the classifiers, namely, EfficienNet, ResNet and DenseNet. In the classification process, the parameter gets optimized by IRV-WSA algorithm, and the parameter tuned are weight in Efficientnet, weight in Resnet and weight in Densenet for increasing of accuracy, NPV, MCC, and also reducing the FNR. At last, this provided the classified leaf diseases as the output by using the high ranging method between the outputs of the classification techniques.

Experimental dataset

The image data is collected in different states namely: healthy and diseased leaf images from the standard Grape vineyards in Coimbatore. The various leaf image dataset is gathered from farm land and web sources. The captured images have the resolution of 1080 × 2412 pixels with random landscape and portrait orientation. This dataset has nearly 1920 images of jpg format respect to 4 different categories. In this case, there are 12 plant diseases are counted. Sample input data of citrus and tomato leaves are shown in the (Fig. 3). Sample input data of grape and apple leaves are shown in the (Fig. 4).The sample input grape plant leaf images with their class labels of healthy and affected are depicted in (Fig. 5). These images are denoted as $sdi_{y}^{in}$. The pre-processing of the input leaf image is crucial for improving the image’s visual impact before further processing. Typically, the gathered photos in the dataset are of such low quality that they need to be sharpened and noise-filtered. A two-dimensional matrix is created from the obtained picture in the dataset during the pre-processing stage, and the RGB image is transformed into a grayscale image. The noise in the leaf image is eliminated by applying a median filter. Increasing the image’s contrast is frequently referred to as enhancement.

Fig. 3

Sample input data of citrus and tomato leaves.

Fig. 4

Sample input data of grape and apple leaves.

Fig. 5

Input data of healthy and affected grape leaves. (a) Healthy leaf, (b) Affected leaf.

Proposed improved random variable-based water strider algorithm (IRV-WSA)

In order to effectively optimize segmentation parameters, the IRV-WSA is an enhanced meta-heuristic algorithm that draws inspiration from the movement and versatility of water striders. Premature convergence is less likely thanks to the adaptive random variables that improve the global search capability. The dynamic adjustment of cluster centroids and membership values maximizes inter-cluster distinction while minimizing intra-cluster variation. By using adaptive mutation and crossover operators, the system improves segmentation accuracy and speeds up convergence. Exploration and exploitation stages are balanced to increase computational efficiency, which lowers processing time without sacrificing accuracy. Subsequently, the AHKM-FCM model incorporates the adjusted parameters from IRV-WSA, guaranteeing improved segmentation performance and resilience across various plant leaf datasets. The IRV-WSA-ETLNet strategy optimizes model performance by combining the enhanced transfer learning (ETLNet) and improved random variable-based water strider algorithm (IRV-WSA). It is better than conventional transfer learning methods in a number of respects. With IRV-WSA-ETLNet, hyperparameters (learning rate, batch size, number of layers, etc.) are automatically adjusted using the IRV-WSA optimization method, improving convergence and model efficiency.

In comparison to conventional transfer learning methods, IRV-WSA-ETLNet performs better by,

• Automatically adjusting hyperparameters for optimal performance.

• Shortening training time and speeding convergence.

• Enhancing generalization and increasing its flexibility in handling novel datasets.

• Improving segmentation quality by feature selection optimization.

• Increasing computing efficiency and adaptability to practical uses.

Advantages of IRV-WSA over PSO & GA algorithms

Greater adaptability

By adjusting its parameters, IRV-WSA may be able to modify exploration and exploitation in real time.

Faster convergence

IRV-WSA could arrive at optimal solutions more quickly than GA’s crossover/mutation and PSO’s velocity updates.

Diminished premature convergence

IRV-WSA may avert stagnation in local optima by enhancing random variable modifications.

The inability of traditional optimizers (PSO, GA, GWO, and WSA) to strike a balance between exploitation (local refinement) and exploration (global search) frequently results in delayed optimization or premature convergence. Other conventional optimizers may not be as effective for complicated, high-dimensional problems since they employ fixed or heuristic-based step sizes. Other meta-heuristics, PSO, and GA need the manual adjustment of factors such as crossover frequency, mutation rate, and inertia weight, which may not translate well to other issues. The IRV-WSA dynamically modifies the ratio of exploration to exploitation in response to search progress. Enables quicker convergence without going over the ideal solution by dynamically adjusting the step size based on the distance between population members. Based on population variety and iteration success, it automatically adjusts settings using an adaptive control technique.

In this developed model, IRV-WSA algorithm is operated to optimize the parameters in both segmentation and classification process. The parameters tuned in segmentation process are the number of attempts in K Means Clustering, parameter and epsilon in Fuzzy C Means and its iterations to increase the variance and correlation coefficient. In the classification process, the tuned parameters are weight in Efficientnet, weight in Resnet and weight in Densenet for increasing of accuracy, NPV, MCC and also reducing the FNR. Here, the reason for selecting the WSA algorithm is it has some advantages like it helps to determine the damaged portion with more accuracy; it is utilized to select the effective features and support to enhance the accuracy of the solution and reliability. However, it has some limitations that are, during mating, the probability attraction is transformed into the form of adaptive instead of static, and the computation time of this algorithm is more. Owing to this issue, we proposed IRV-WSA for providing the best outcome. In developed IRV-WSA, the adaptive concept is used to determine the random numbers $r_{1}$ and $r_{2}$, and it is explained in Eq. (1) and Eq. (2).

$$r_{1} = \frac{{\left( {bst\,fit * bst\,solu} \right)}}{{\left( {me\,fit * me\,solu} \right)}}$$

(1)

$$r_{2} = \frac{{rd\left( {0,1} \right)}}{{r_{1}^{2} }}$$

(2)

In Eq. (1), Eq. (2), the term $r_{1}$ and $r_{2}$ represents the random numbers in the developed IRV-WSA algorithm. The terms $bst\,fit$,$bst\,solu$,$me\,fit$ and $me\,solu$ represents the best fitness, best solution, mean fitness and mean solution, respectively.

WSA³⁰

The algorithm of WSA has five important stages that include births, territory establishments, death, mating, and feeding. Through the WSA algorithm, search spaces are termed as the lake containing various solutions for territories and serving food as the metaphors for objective functions.

Initial birth: water striders (WSs) are born by the eggs dispersed in lakes. The random dispersion is determined by using Eq. (3).

$$A_{j} = Tc + rd(Tc - Kc),j = 1,2,...,oxt$$

(3)

In Eq. (3), the term $A_{j}$ describes the initial positions of $j^{th}$ water striders. The term $Tc$ and $Kc$ represents the bottommost and topmost bounds corresponding to the minimum and maximum value of allowable, $rd$ denotes the erratic numbers in the range from 0 to 1 and $oxt$ denote WS’s number. The initial WS is calculated with the help of objective functions that evaluate the fitness position on lake.

Territory establishments: The WSs maintain the territories for living, mating and feeding. The upcoming method is utilized to assign the WSs in territories to develop the territory number. Initially, WSs are selected by basing the fitness, and they are sectioned into $\frac{oxt}{{ou}}$ classes in order. The members $j^{th}$ in every group are assigned to the territory of $j^{th}$, and $j = 1,2,....,ou$. Thus total $\frac{oxt}{{ou}}$ number WSs will live in every territory. The feminine are generally chosen the better location in every territory to feed. The worst and best fitness of position in every territory are taken as male and female respectively.

Mating: It is a remarkable development in WSs life. The male sends courtships calling ripple and females respond for sending the attractions. The probability to send the response for attraction is assumed equivalent to $q$, and hence $(1 - q)$ probability which is gone from repulsive responses. So, the responsibility of feminine isn’t identified; it is considered as 50% of $q$. If the feminine transmits the fascinate signals, they forward towards each and get a reproduction. After reproduction, the new positions are upgraded in terms of location among them. If the feminine avoids the requests, the keystone (male) mounts the female, and next the dismounts the male. The male may be in state of not mate or in mate, the newly updated position for the male is determined using Eq. (4).

$$\left\{ \begin{gathered} A_{j}^{u + 1} = A_{j}^{u} + Q.r_{1} \,(if\,mate\,occurs\,with\,probility\,o) \hfill \\ \,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,otherwise \hfill \\ A_{j}^{u + 1} = A_{j}^{u} + Q.(1 + r_{2} ) \hfill \\ \end{gathered} \right.$$

(4)

Here, the term $A_{j}^{u}$ denotes $j^{th}$ position of WS in $u^{th}$ iteration, $r_{1}$ and $r_{2}$ describes the random value among [0, 1] in the conventional algorithm, whereas in the developed IRV-WSA, it is determined with the adaptive concept using the Eq. (1) and Eq. (2). The term $Q$ represents the vector value, and the initial and the final point’s positions of male and female are $A_{j}^{u - 1}$ and $A_{E}^{u - 1}$_.

This feminine is chosen by the fitness mechanism of roulette wheel selections. Equation (5) describes the Euclidean distances among females and males.

$$Q = A_{E}^{u - 1} - A_{j}^{u - 1}$$

(5)

Here, the term $Q$ denotes the length and it is equivalent to Euclidean distances among female and male.

Feeding: The method needs more energy, if the mating is successfully happened or not. Hence, in newly updated location, WSs hunt food. The food available to the location is calculated by using objective functions. If the objective function’s value is more than preceding state, then they found the food and in else condition, it needs to move more to get the high fitness. The generation of new position is elaborated in Eq. (6).

$$A_{j}^{u + 1} = A_{j}^{u} + 2rd.(A_{AK}^{u} - A_{j}^{u} )$$

(6)

Equation (6) explains the transfer to a newly updated position near the better WS in lake contains good resources for food.

Death and succession: To evaluate the conclusion for the process of food attainments, there has to evaluate the objective functions, and then it compares with superior positions. The WSs get to die, if the newly obtained fitness value is less, since it only not discovers the food and, it also enlarges the danger battle to the WS of destination territories. Here, the matured larvae have succeeded the death as male, and his position is placed within the territories is represented in Eq. (7).

$$A_{j}^{u + 1} = KA_{j}^{u} + 2rd.(Tc_{k}^{u} - KA_{k}^{u} )$$

(7)

In Eq. (7), the term $KA_{k}^{u}$ and $Tc_{k}^{u}$ represents the minimum and maximum value of the position of WS within $k^{th}$ a territory. And also, they define the boundary for the territory of WSs died. The pseudo-code for the developed IRV-WSA approach is stated in Code 1, and the data flow for the developed IRV-WSA method is described in (Fig. 6).

Fig. 6

Flowchart for the developed IRV-WSA algorithm.

Adaptive hybrid K-means with fuzzy C-means clustering for segmentation (AHKMC-FCM)

K-means clustering (KMC)

KMC³¹ actively chooses the category numbers and its calculations of the closeness on Euclidean distances among the point. The pictorial illustration for the KMC model is provided in (Fig. 7).

Fig. 7

Pictorial illustration of KMC architecture.

In this algorithm, the clusters $L$ are $B = \left\{ {B_{1} ,B_{2} ,...B_{L} } \right\}$ arbitrarily chosen the partition and $m$ the sample datasets $C = \left\{ {y_{1} ,y_{2} ,....y_{m} } \right\},m \ge k$ sectioned into near clusters. Next, re-evaluate the cluster’s middle point. Then, the process is repeated till the condition of convergence is attained. The function of convergence is expressed in Eq. (8).

$$A(B) = \sum\limits_{L = 1}^{L} {\sum\limits_{{y_{j} \in B_{L} }} {\left\| {y_{j} - B_{L} } \right\|} }^{2}$$

(8)

In Eq. (8), the term $A$ represents the least square error (SSE) of a cluster; the term $B_{L}$ denotes the cluster’s middle point. The deduced and verified is the arithmetical meaning for mid point. The result proved that the best midpoint of the cluster is termed to be a mean value in every point in the cluster. The mean value is computed in Eq. (9).

$$B_{L} = \frac{{\sum {_{{y_{j} \in B_{L} }} y_{j} } }}{{\left| {B_{L} } \right|}}$$

(9)

Equation (9) shows the cluster’s closeness samples nearest to the middle. The scope of KMC is to get less clustering results. When $A$ is smaller, it helps to increase the clustering degree. This is the easy method to fasten the speed, good scale and generate the middle point.

Fuzzy C-means clustering (FCM)

FCM³² reduces the objective functions that evaluate the total number of squared distances from every center of the cluster. It changes between evaluating the centre of clusters for giving the value of membership of every data sample and evaluating the value of membership in the clusters. The objective function is denoted as $I$ , and it is described in Eq. (10).

$$I(T,U) = \sum\limits_{j = 1}^{b} {\sum\limits_{l = 1}^{m} {v_{jl}^{n} } } C_{jl} (y_{l} ,w_{j} )$$

(10)

In Eq. (10), the term $m$ describes the sample numbers, $n$ describes the fuzzifier, $b$ expresses the cluster numbers, $T$ denotes the matrix for membership, and $v_{jl}$ defines the value of membership for $l^{th}$ the data sample in $j^{th}$ cluster. The term $U$ describes the centre of cluster, and the term $w_{j}$ defines the centre of $j^{th}$ cluster. The functions for determining the value of membership and centre of the clusters are illustrated in Eq. (11) and Eq. (12).

$$v_{jl} = \frac{{C_{jl} (y_{l} ,w_{j} )^{{\frac{1}{1 - n}}} }}{{\sum\limits_{k = 1}^{d} {C_{jl} (y_{l} ,w_{j} )^{{\frac{1}{1 - n}}} } }}$$

(11)

$$w_{j} = \frac{{\sum\limits_{k = 1}^{m} {C_{jl} (v_{jk} )^{{^{n} }} y_{k} } }}{{\sum\limits_{k = 1}^{m} {(v_{jk} )^{n} } }}$$

(12)

In Eq. (11) and Eq. (12), the term $D$ denotes the squared distances between $l^{th}$ data sample and the centre of $j^{th}$ cluster. There is various ways to terminate and initialize the algorithms. Any applicable set value may be utilized to initialize the value of $T$ or $U$ matrix. Typically, the center of cluster $U$ is worked by selecting $b$ samples from every datasets. This KMC algorithm gets terminated when a variation among the evaluated matrix norm for a successful membership matrix doesn’t exceed a user’s provided parameter $\varepsilon$. The graphical explanation for the FCM framework is shown in (Fig. 8).

Fig. 8

Diagrammatic representaion of FCM model.

Developed AHKM-FCM-aided plant leaf disease segmentation

To improve segmentation accuracy and resilience, the suggested AHKM-FCM segmentation approach combines the advantages of K-Means and Fuzzy C-Means clustering. The reason for choosing these algorithms is they have significant features. The advantages of KMC are providing better group formation of same data types and adapting with the new environment. The advantages of FCM are providing better outcome and working harder than other clustering models. For all these significant features, we select these algorithms to develop the AHKM-FCM model for improving the segmentation task. Because AHKM-FCM is hybrid, it is very good at detecting accurate and smooth leaf borders. This holds particular significance in the segmentation of plant leaves, as determining the precise leaf contour is essential for further research (e.g., calculating leaf area, diagnosing diseases). The leaf areas are initially crudely segmented using K-Means. Using the Elbow Method, the number of clusters is dynamically determined. The output clusters are used as the starting point for the subsequent phase. The K-Means divided zones are further refined with the help of fuzzy C-means (FCM). With FCM, boundary precision is improved by allocating pixels to many clusters with varying membership values. The segmentation results are refined by repeated minimization of the objective function.

Although K-means’ frequent centroid updates can make it computationally costly, its hybridization with FCM offers a quicker convergence rate without compromising segmentation quality, making it a viable option for large-scale datasets of plant leaf images. In segmentation stage, the sample data $sdi_{y}^{in}$ are fed as input. After that, parameters are optimized in the developed AHKM-FCM technique by using IRV-WSA algorithm. The tuned parameters are attempt numbers in KMC, fuzziness parameters in FCM, epsilon in FCM and occurrence count in FCM for improving the variance and correlation coefficient. After completing this process, it provides the segmented images as the output. The segmented image output is represented by $SE_{x}^{seg}$. The fitness mechanism for the proposed AHKM-FCM model is explained in Eq. (13).

$$fn_{1} = \mathop {\arg \,\min }\limits_{{\left\{ {atmt_{t}^{Seg} ,fuzz_{t}^{seg} ,epsi_{t}^{seg} ,iter_{t}^{seg} } \right\}}} \,\left( {\frac{1}{Va + cor\,}} \right)$$

(13)

In Eq. (13), the terms, $fn_{1}$ represents the fitness,$Va$ denotes the variance and, $cor$ describes the correlation co-efficient. The terms, $atmt_{t}^{Seg}$ represents attempts number in KMC at the range $\left[ {2,10} \right]$, $fuzz_{t}^{seg}$ denotes fuzziness parameters in FCM of range $\left[ {2,20} \right]$, $epsi_{t}^{seg}$ describes epsilons in FCM of range $\left[ {1,10} \right]$ and $iter_{t}^{seg}$ denotes iterations numbers in FCM of range $\left[ {10,100} \right]$. The representation of variance and correlation coefficient are described in Eq. (14) and Eq. (15).

$$va = \sum\limits_{j = 1}^{m} {(SE_{x}^{seg} - \overline{{SE_{x}^{seg} }} )^{2} }$$

(14)

$$Cc = \frac{{T_{ab} }}{{T_{a} T_{b} }}$$

(15)

In Eq. (14) and Eq. (15), the term $SE_{x}^{seg}$ denotes the segmented images and $\overline{{SE_{x}^{seg} }}$ represents the mean values of the segmented images. And, the term $T_{ab}$ describes the sample covariance and $T_{a} T_{b}$ denotes the sample standard deviation. The pictorial illustration of AHKM-FCM- model for leaf disease segmentation model is detailed in (Fig. 9).

Fig. 9

Pictorial illustration of AHKM-FCM approach.

Using IoU (Intersection over Union) and other pertinent metrics, segmentation quality should be quantitatively assessed in leaf disease prediction to guarantee precise identification of sick areas. Analysis of Segmentation Quality results: Pixel accuracy: 0.875; IoU (Intersection over Union): 0.778; dice coefficient: 0.875. According to these findings, the predicted mask exhibits great segmentation accuracy and a good overlap with the ground truth.

By combining the advantages of K-Means and fuzzy C-Means (FCM), the segmentation quality is eventually improved by overcoming the drawbacks of each method alone. By providing both quick convergence and excellent segmentation quality, this hybrid technique greatly enhances performance in plant leaf segmentation. In comparison to conventional segmentation techniques, AHKM-FCM offers a reliable and flexible method for plant leaf segmentation that offers faster convergence, improved border identification, and the ability to handle noisy data.

Parameter optimizer using deep learning

EfficientNet, ResNet, and DenseNet were chosen due to their distinct benefits in the categorization of plant diseases is explained in the (Table 1).

Table 1 Comparison of EfficientNet, ResNet, and DenseNet.

EfficientNet

The efficientnet³³ is introduced as incredulous technique based on neural networks for improving the resolution, breadth and depth with the set of preset parameter, contrasting the standard practices that randomly adjust all these factors. This has 18 number of convolution layers $C = 18$, and every layers are provided with kernels $k(3,3)$ or $k(5,5)$.

The image is given as input that has three channel colors of red, green and blue, all these are corresponding to $224 \times 224$ sizes. The layer placed after is scaled downward in the resolution for reducing the size of feature maps, but scaling upward in the width improved the accuracy value. For any instances, second layer in convolution contains $V = 16$ number of filters, and the convolution of the next layer has $V = 24$ numbers of filter. The maximum filter number is $C = 1$, $280$ for the final layer that feds to fully connected layer. The graphical discrimination for the EfficientNet model is explained in (Fig. 10).

Fig. 10

Graphical explanation for EfficientNet framework.

Comparing IRV-WSA-ETLNet’s performance against that of cutting-edge transfer learning models like EfficientNet (without optimization) is crucial to highlighting its benefits. EfficientNet is among the most potent deep learning models for segmenting and classifying images. But instead of using metaheuristic optimization, it depends on default hyperparameters. Through comparison:

• (i) EfficientNet (not optimized but pre-trained).

• (ii) Transfer learning with standard fine-tuning (ETLNet).

• (iii) IRV-WSA-ETLNet: IRV-WSA-optimized transfer learning.

Anticipated result: IRV-WSA-ETLNet should provide better accuracy and faster convergence than EfficientNet (without optimization) because of adaptive hyperparameter adjustment.

ResNet

There are several types of neural network methods developed based on deep learning techniques, but the training challenges in CNN method gets increases as before. To solve this difficulty in deep learning based-CNN model, Kai Zhang and team developed the model named as ResNet. The model of ResNet³⁴ more decreases the training difficulty in deep neural networks by developing the mapping identity regarding the CNN model. They reduce the propagations of the error and tune the model parameter. Moreover, it also attained the better result in computed vision-based approaches namely, segmenting the images, targets positioning, recognizing the images, etc.

Basically, the model of ResNet³⁵ is CNN’s updated version, which is termed as the core model utilized to identify the errors in wind turbines. The illustration of the ResNet model is explained in the (Fig. 11). It’s normally developed for significant basic block, including the input layers, order of the convolutional layer, Batch Normalizations (BN), mapping identity, and fully connected layers of output. As same as the model of standard convolutional, ResNet too has the fundamental residual block. It’s the common residual blocks, in which a key processing path of the model is “BN $\to$ ReLU $\to$ Convol $\to$ BN $\to$ ReLU $\to$ Convol” and finally added to cross layers paths (mapping identity) for developing the entire basic model of residual blocks.

Fig. 11

Graphical representation for ResNet framework.

DenseNet

DenseNet³⁶ is the CNN’s type that is proposed in effective and efficiently in recognizing the images. The main innovation of the DenseNet is the dense connections with every layer proving the inputs from the entire preceding layer. This facilitates more proficiency outcome and learning more than image identification approaches of other CNN. The advantages of DenseNet are that it reduces the parameters and computations and identifies the challenges in vanishing gradients. The dense connection is termed incessantly connecting a preceding layer of the feature maps as the inputs to the next layers that strengthen the information flows among the levels. The DenseNet201³⁷ model has a transitions layer and dense blocks. The dense blocks contain growth rate, and multiple layer and bottleneck layer is connected by using the concatenation based on the channels. This process improves the parameter f the network and also reduces the computing efficiency.

To solve this increasing parameter issue, it identifies the result that is utilizing the growths rate $( = l)$ as the hyper-parameters, next apply the “BN $\to$ ReLU $\to$ convolution (1 × 1)$\to$ nonlinear transformations $\to$ DenseNet model”. Moreover, it’s used for dropping the counts of an input features map and improved the efficiency of calculation. The layers of transition have the responsibility to lower the height and width of each feature map and its numbers. It is connected after the model of dense block and comprised as “BN $\to$ ReLU $\to$ convolution $(1 \times 1)$ $\to$ average pooling”. Now, this factor lies among 0 and 1, and it is utilized to identify how many feature maps are reducing the sizes. The diagram for DenseNet model is shown in (Fig. 12).

Fig. 12

Diagrammatic representation for DenseNet framework.

Developed ETLNet for plant leaf disease classification

The segmented images $SE_{x}^{seg}$ are considered as the input data in this developed model. ETLNet model is used to group the leaf disease which is present in the plants. It comprises the classifiers such as EfficientNet, ResNet and DenseNet architecture. The parameters are tuned in this developed ETLNet model by IRV-WSA algorithm for enhancing the classification operation. The tuned parameters are weight in Efficientnet, weight in Resnet and weight in Densenet for increasing accuracy, NPV, MCC, and reducing the FNR. The final classified output is provided by using a high-ranking approach. The high ranking method is nothing but the final value attained by majority of classified value acquired from DensNet, EfficientNet and ResNet. The classified image output is obtained as the identified leaf diseases of the plant images. The fitness function for the ETLNet classification model is described in Eq. (16).

$$fn_{2} = \mathop {\arg \,\min }\limits_{{\left\{ {we_{t}^{clas} ,wr_{t}^{clas} ,wd_{t}^{clas} } \right\}}} \,\left( {\frac{1}{AC + NP + MC\,} + FN} \right)$$

(16)

Here, the terms $fn_{2}$ represent the fitness,$AC$ denotes the accuracy and $NP$ describes the NPV. The term $MC$ describes the MCC and $FN$ denotes the FNR. The term $we_{t}^{clas}$ represents weights in EfficientNet of range $\left[ {0,1} \right]$, $wr_{t}^{clas}$ denotes weights in ResNet of range $\left[ {0,1} \right]$, $wd_{t}^{clas}$ and describes weights in DenseNet of range $\left[ {0,1} \right]$ respectively.

Result and discussions

Experimental setup

The designed framework was replicated executed on Python. The designed algorithm had the 10 numbers of populations and also the total iteration of 50. The recommended IRV-WSA-WTLNet model was explored and executed with several models for demonstrating the precision and accuracy of the framework. The recommended framework was evaluated with several algorithms such as, Mine Blast Optimization (MBO)-ETLNet, Chameleon Swarm Optimization (CSO)-ETLNet, Archimedes Optimization Algorithm (AOA)-ETLNet and Water Strider Algorithm (WSA)-ETLNet³⁰. Moreover, the classifiers are Efficientnet³⁰, ResNet³¹, DeneNet³³ and Ensemble Network. The details of implementation are listed as below.

• Software used—PyCharm—version 3.1.1, Anconda—Version 3 (https://www.jetbrains.com/pycharm/download/other.html)³⁸

• Programming Language—Python

• Packages—> TensorFlow, Keras, OpenCV-Python, prettytable, numpy

Performance measures

The subsequent measures investigate the operation of the proposed model. The formulation of MCC, FPR, FNR, FDR, and NPV are presented below.

True positive (TP)

The objective anomaly was correctly discovered and is there, according to the test result.

True negative (TN)

The test output shows that the objective anomaly was not detected and did not exist.

False positive (FP)

The test’s outcome shows that the objective anomaly was correctly identified and did not exist.

False negative (FN)

The test result shows that the anomaly was not correctly diagnosed and that there is no objective abnormality.

Matthews’s correlation co-efficient (MCC)

It is a fair measure for assessing classification models, especially in cases when the datasets are unbalanced. It is measured by the Eq. (17),

$$\text{MCC}=\frac{\left(TP * TN\right)-(FP * FN)}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)(TN+FN)}}$$

(17)

False positive rate (FPR)

It is a statistic that calculates the percentage of real negative situations that are mistakenly labeled as positive. It is calculated by the Eq. (18),

$$\text{FPR}=\frac{FP}{FP+TN}$$

(18)

False negative rate (FNR)

It calculates the frequency with which real positive situations are mistakenly labeled as negative. It is measured by the Eq. (19),

$$\text{FNR}=\frac{FN}{FN+TP}$$

(19)

False delivery rate (FDR)

It is used to calculate the percentage of positive instances among all successful deliveries. It is evaluated by the Eq. (20),

$$\text{FDR}=\frac{FP}{FP+TP}$$

(20)

Negative predictive value (NPV)

It is a statistic that quantifies the likelihood that a negative prediction is true in classification tasks. It is measured by the Eq. (21),

$$\text{NPV}=\frac{TN}{TN+FN}$$

(21)

Accuracy

The accuracy of the model, an important performance metric, is used to test our presumptions on the negative and ideal classes. Accuracy is identified using Eq. (22).

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

(22)

Sensitivity

By simply computing the ratio of all Positive aspects of everything favorable projections, one can get the sensitivity. Sensitivity is calculated by utilizing Eq. (23).

$$\text{Sensitivity}=\frac{TP}{TP+FN}$$

(23)

Specificity

Indicators of specificity are the proportion of real numbers has the potential to precisely match each actual number. It is evaluated using Eq. (24),

$$\text{Specificity}=\frac{TN}{TN+FP}$$

(24)

Precision

The degree of precision shows how well the different measurements agree with each other. It predicts using the Eq. (25).

$$\text{Precision}=\frac{TP}{TP+FP}$$

(25)

F1 score

It strikes a compromise between memory and accuracy. It predicts using the Eq. (26).

$$\text{F}1\text{ Score}=\frac{2*recall*precision}{recall+precision}$$

(26)

Developed AHKM-FCM-based segmented results

The pre-processed leaf image is fed into the segmentation stage. The segmented result images of the affected and non-affected part of given input leaf image is depicted in the (Fig. 13).

Fig. 13

Segmented result of healthy part and affected part of input leaf image.

The image description of five leaf images and its corresponding segmented output images of the proposed AHKM-FCM segmentation model are explained in (Fig. 14).

Fig. 14

Segmented result of developed AHKM-FCM-based plant leaf disease segmentation framework.

Confusion matrix analysis on developed model

The confusion matrix analysis of the developed model is described in (Fig. 15). It refers to as a table that aids to estimate the accuracy value for classifying the disease. It is calculated with “true positive, false positive, true and false negative”. This matrix also represents with the comparison of predicted value of target and actual value of classified result. The label represents the different disease classes of plant leaves that are given as “0- Healthy, 1-Bacterial_spot, 2-Early_blight and 3-Late_blight”. Hence, for each and every class, the true and false values are determined. With the help of this matrix, the performance metrics are computed to ensure the system efficiency. It is used to give the prediction analysis in the form of matrix.

Fig. 15

Confusion matrix framework.

Proposed model cost function analysis (CFA)

The CFA of the developed model is elaborated in (Fig. 16). The cost function analysis is helping to evince the high convergence rate of proposed optimization algorithm. It is evaluated by varying the number of iteration, in which the cost value for proposed work attains less. Hence, the less cost function leads to maximize the convergence rate while providing the optimum results. The cost function of the developed IRV-WSA-ETLNet framework is 8.33% enhanced than MBO-ETLNet, 15.38% more than CSO-ETLNet, 21.43% improved than AOA-ETLNet and 21.49% superior than WSA-ETLNet at 10th iteration. This demonstrates the efficiency of the IRV-WSA-ETLNet model that is compared with other conventional techniques. IRV-WSA has a fixed step-size exploration, whereas regular WSA dynamically adjusts parameters dependent on the optimization landscape. Premature convergence is avoided and subsequent solution fine-tuning is improved by this flexibility.

Fig. 16

Cost function analysis framework.

Algorithmic-based analysis on developed model

The algorithmic analysis of the suggested model is represented in (Fig. 17). Here, the comparison is taken place with proposed and traditional heuristic algorithms for ensuring the system efficiency. The main intent of proposed IRV-WSA is to render the best values for the parameters, which further helps to group the plant disease. In Fig. 16d, the FPR of the proposed framework is, 63.08% more than MBO-ETLNet, 56.36% improved than CSO-ETLNet, 51.02% increased than AOA-ETLNet and 37.66% superior than WSA-ETLNet when the value of epoch is at 500. Thus, it proves the designed framework has best performances.

Fig. 17

Algorithmic analysis framework concerning (a) Accuracy, (b) FDR, (c) FNR, (d) FPR and (e) MCC.

Classifier-based analysis on developed model

The classifier analysis model is described in (Fig. 18). In Fig. 18a, the accuracy of the designed IRV-WSA-ETLNet framework is 5.88% improved than EfficientNet, 4.65% greater than ResNet, 3.45% enhanced than DenseNet and 2.27% more than Ensemble Network when the epoch is at 400. By varying the number of epochs, the performance evaluation is done and compared with other existing deep learning models (without optimizing the parameters). This can be analyzed with true and false measures. From the findings, the true rate is maximized and false value is reduced. This showed that the model’s performance has better accuracy than other existing approaches.

Fig. 18

Classifier analysis framework concerning (a) Accuracy, (b) FDR, (c) FNR, (d) FPR and (e) MCC.

ROC-based analysis model

The ROC analysis of the model is described in (Fig. 19). In this figure, the true positive rate of the recommended framework is, 10.71% more than EfficientNet, 8.14% increased than ResNet, 5.68% more than DenseNet and 3.33% enhanced than Ensemble network when the false positive rate at 0.4. Therefore, the result of the ROC analysis demonstrated that model is superior to other available approaches.

Fig. 19

ROC analysis framework.

Overall algorithmic analysis model

The overall algorithmic comparison analysis model is described in (Table 2). The overall assessment is achieved for existing and proposed optimization algorithm; where it delivers the numerical value of multiple measures when consider the epoch count of network as 600. In Table 2, the NPV of the developed method is 7.17% more than MBO-ETLNet, 5.31% improved than CSO-ETLNet, 3.74% increased than AOA-ETLNet and 2.04% superior to WSA-ETLNet. The findings of various existing models analyzed from^39,40, and⁴¹.

Table 2 Overall algorithmic analysis framework.

The overall algorithmic analysis from Table 2, the corresponding graph was plotted in (Fig. 20). The suggested IRV-WSA-ETLNet model achieves the accuracy of 94.853%, sensitivity of 94.750%, specificity of 94.888%, precision of 96.068%, and F1 Score of 90.201%. Comparing with existing models our proposed model has more efficiency.

Fig. 20

Algorithmic analysis graph.

Overall classifier analysis model

The overall classifier comparison of proposed with existing model analysis is described in (Table 3). In this analysis, the traditional classifier models are taken to prove the interpretation of suggested classifier network in terms of true and false measures. In Table 3, the MCC of the suggested model is 12.73% increased than EfficientNet⁴², 10.93% more than ResNet⁴³, 6.51% superior to DenseNet³⁷ and 1.90% more than Ensemble network⁴⁴. This confirms the proposed IRV-WSA-ETLNet has more performances in every metrics.

Table 3 Overall classifier analysis framework.

The comparable graph for the entire classifier analysis from Table 3 was presented in (Fig. 21). The accuracy, sensitivity, specificity, and F1 Score of the proposed IRV-WSA-ETLNet model are 94.853, 94.750, 94.888, and 96.068%, respectively. Our suggested system is more efficient than the current systems.

Fig. 21

Classifier analysis graph.

Statistical analysis of the model

The overall statistical analysis of comparison of the traditional and suggested model is explained in (Table 4), the mean value of the proposed framework is 0.005% improved than MBO-ETLNet, 0.317% more than CSO-ETLNet, 0.330% more than AOA-ETLNet and 0.289% enhanced than WSA-ETLNet. This describes the developed IRV-WSA-ETLNet model has more performance when compared with different current heuristic models^39,40, and⁴¹.

Table 4 Statistical analysis framework.

Comparing the Baseline CNN, ETLNet, and IRV-WSA-ETLNet allows us to show, why accuracy increases and training speeds up with transfer learning. How IRV-WSA enhances performance through transfer learning optimization. Because baseline CNN lacks pre-trained information, it performs badly. Using pre-trained features, ETLNet (Transfer Learning) increases speed and accuracy. For maximum accuracy and efficiency, IRV-WSA-ETLNet further enhances transfer learning.

A five-fold k-fold cross-validation strategy was used to conduct cross-validation in order to better guarantee the model’s dependability. This assisted in assessing the consistency of the model and minimizing overfitting by testing and training on various dataset subsets. Real-world deployment or field testing to evaluate a leaf disease prediction models performance outside of lab settings covered in the paper for a comprehensive assessment. Real-world deployment was carried out in agricultural settings to verify the suggested leaf disease prediction model’s practical efficacy. Farmers may now take pictures of leaves with their smartphones thanks to a mobile application that incorporates the model. Field tests were also conducted in a number of places with various environmental factors, such as variable lighting, backdrops, and leaf orientations. In order to evaluate the robustness of the model, important performance measures including accuracy and intersection over union (IoU) were calculated and contrasted with controlled laboratory settings.

Uncertainty is measured using confidence intervals (CI) and p-values from statistical tests (such as the t-test). In the paired t-test, the significant differences for IoU (t = 13.73, p = 0.005) and Dice (t = 14.86, p = 0.0045) are as follows: Dice: (0.00386, 0.00394) and IoU: (0.969, 1.007) in 95% Confidence Intervals. The boxplots that illustrate the distributions of IoU and Dice scores for various segmentation techniques are shown below. Stability of IoU Score: While the suggested deep learning model has a little greater variance, suggesting adaptation to complicated scenarios, the conventional approaches (Otsu, Watershed, GrabCut) demonstrate strong IoU consistency (narrow variance). Dice Score Stability: Across all approaches, dice scores exhibit very little variation and stay very close. Computational Time duration: How long anything lasts from start to finish is known as its entire time length. The overall computational time analysis of comparison of the traditional and suggested model is explained in (Table 5) and its corresponding graph were shown in the (Fig. 22). The MBO-ETLNet takes 42.365 ms, CSO-ETLNet takes 39.246 ms, AOA-ETLNet consumes 34.657 ms, and WSA-ETLNet takes 29.432 ms. And the newly constructed IRV-WSA-ETLNet model consumes less computational time duration of 24.378 ms, compared with existing models.

Table 5 computational time.

Fig. 22

Computational time analysis graph.

The following hyperparameters were used for training the model: Learning Rate: 0.001, Batch Size: 32, Adam as the Optimizer, 50 Epochs, and categorical cross-entropy as the loss function, the proposed system loss function decreasing validation shown in (Fig. 23). Using this model in conjunction with conventional approaches (such as Thresholding, Watershed, or GrabCut), it may visually compare the segmentation results against industry standards. Constructing an implementation in Python, that are, loads a leaf picture sample along with its segmentation mask of ground truth. Use classic segmentation methods (GrabCut, Watershed, and Otsu’s Thresholding). Relates them to the forecast of a deep learning model, IoU scores are calculated for every procedure. A paired t-test on the IoU and Dice scores used to evaluate the segmentation findings’ statistical significance. The suggested deep learning model performs noticeably better than conventional techniques, it was determined. The average and standard deviation of each method’s IoU and dice scores.

Fig. 23

Proposed model loss curve.

Conclusion

This work has demonstrated the efficient plant leaf detection model using intelligent-based techniques with the help of leaf images of various plants. To conduct the process, raw input images were collected from publicly available online sources. Next, the abnormal regions of the leaves were segmented using the proposed AKHM-FCM method. AKHM-FCM, the parameters were selected optimally by IRV-WSA approach. This novel approach improved segmentation accuracy. Lastly, the resultant segmented images were fed unto ETLNet, which has influenced the functionality of EfficientNet architecture, ResNet and DenseNetmodels for classifying the leaf diseases. To reduce complexity, hyper-parameter tuning was performed using IRV-WSA, which improved performance. The suggested IRV-WSA-ETLNet model has the following classification outcome results: 94.853% accuracy, 94.750% sensitivity, 94.888% specificity, and 96.068% F1 Score, and this system consumes less computational time duration of 24.378 ms. The best value of the developed model was 10.78% enhanced than MBO-ETLNet, 0.58% more than CSO-ETLNet, 15.53% superior than AOA-ETLNet and 6.29% enhanced than WSA-ETLNet. Hence, the result proved that the recommended model outperformed than other techniques. Since real-time images are used for segmentation, noise and artifacts may reduce image quality. Various pre-processing techniques are recommended to remove blur and other artifacts these issues will be resolved in further studies or implementation. Farmers may include the suggested plant disease detection model into their regular farming operations by utilizing IoT devices, and mobile applications to optimize its advantages. By integrating an AI-driven disease detection system into their operations, farmers can guarantee both food security and financial stability, which will increase production, decrease losses, and improve agricultural management. By using drone-based monitoring, high-resolution cameras on drones may improve efficiency and save human labor by capturing large-scale agricultural photos in real-time. Integrating IoT sensors to capture leaf images periodically could further enhance the proposed system for precision farming. Farmers will benefit from these advancements by being able to lower expenses, boost output, and guarantee sustainable agriculture.