CSIU-Net+: Pepper and corn leaves classification and severity identification using hybrid optimization

An agricultural product plays a major role in the economical growth of developing countries. Agricultural products like pepper and corn are the essential crops with respect to human health food security. But, these two crops are prone to different diseases such as gray leaf spot, common rust and fruit rot which affects the productivity of crops. Further, the severity identification is also a challenging one. To address these limitations, this work presents different approaches for identifying the crop lesions and predicting the severity and thereby increasing the productivity of crops. The development of the proposed model includes steps such as dataset collection, noise removal, segmentation, feature extraction, classification and severity prediction. Initially, the crop images are pre-processed by the median filter and the pre-processed images are processed are segmented, extracted and classified by the optimized U-Net model. Moreover, hybrid optimizer which is the integration of GJA (Golden jackal algorithm) and RDA (Red deer algorithm) are utilized for precise segmentation and classification. Finally, the severity prediction is computed for the diseased leaves by the measuring the size of image pixels. The experimentation is carried out on the benchmark PlantVillage dataset; the accuracy and precision values achieved are 99.2% and 99.1%. Thus, the experimental outcomes show the effectiveness and stability of the model.


Introduction
In the past centuries, plant diseases have caused prominent disease outbreaks resulting in famine and death on a large scale [1].Because of the enormous amount of crop diseases, crop productivity is reduced where the affected area is not perceived by the naked eye [2].However, most crop diseases are visible, and an experienced agronomist diagnoses them via the optical observance of diseased plant leaves [3].However, climatic variations and the spread of diseases in nearby areas are to be detected at the earlier stage to prevent huge losses in the food grains.Some of the common food crops like rice, corn, pepper, millet, etc are highly affected by various diseases that cause huge productivity losses to the farmer [4,5].Of this, pepper and corn crops are highly affected by bacterial, fungal, and viral attacks resulting in huge ecological and economic losses.
Pepper is one of the primitive crops that provide highly nutritious to human beings thereby reducing several harmful diseases effectively [6].Pepper crops are highly grown by farmers worldwide because of increased consumer demand.Moreover, the corn plant which is scientifically termed Zea mays cultivated on more than 250 million hectares across the globe.Unlike other crops, pepper, as well as corn crops, are also affected by various attacks that directly influence crop productivity [7,8].Especially, in corn crops various diseases like gray leaf spots, northern leaf blight, and spots that form oblique, elongated, tan-brown cigarshaped, and circular lesions around the leaf margins [9].These types of disease attacks lead to financial losses to the farmers and hence effective techniques are necessary to detect pepper and corn leaf diseases accurately.To overcome this issue, farmers utilize pesticides to control the spread of diseases but it is optimal only when detected at the earlier stage [10].People in many countries highly rely on these crop varieties and highly effective techniques are required to control and prevent disease attacks [11,12].
To tackle this issue, an effective and automatic technique is highly necessary to detect pepper and corn leaf diseases with negligible time complexity.Recently, artificial intelligence (AI) based machine learning (ML) and deep learning (DL) techniques have played a vital part in the agricultural field, especially in the disease detection process [13,14].Recently, ML-based techniques like k-nearest neighbor (KNN), multilayer perceptron (MLP), decision tree (DT), artificial neural network (ANN), and random forest (RF), and have been introduced in several studies to recognize crop leaf diseases effectually [15].Despite the current practices performing well, there are limitations like low precision, high computational complexity, and deprived generalization ability.Moreover, the existing ML techniques increase the training intricacy while handling larger databases [16].
Currently, the DL techniques are utilized by several research studies and have proven to have outstanding performance in detecting crop diseases with minimal features.In the agricultural field, various DL systems like long short-term memory (LSTM) and convolutional neural networks (CNNs) have been introduced for the identification of diseases affecting crop productivity [17].It is proven that DL studies become the most promising resolution for detecting crop diseases efficiently.However, obtaining improved classification performance necessitates an increase in the complexity of the network model thereby maximizing the training fluctuations [18,19].Hence, a lightweight optimized network model is introduced in the proposed study to classify the pepper and corn plant leaf diseases at less time complexity.This work aims to address the shortcomings in prior research by introducing an innovative optimized UNet model based semantic segmentation.The objective is to enhance the plant disease management system robustness through lesions classification and severity estimation.

Motivation
Recently, crop diseases have been the major causes that directly influence agricultural productivity.Of this, pepper and corn crops are highly affected by bacterial spot, common rust, and gray leaf spot disease that causes huge economic losses to the farmers.Identifying the pepper and corn leaf disease earlier may reduce agricultural damage and crop efficiency.Conventional manual detection techniques require expertized Agronomists leading to increased time consumption and cost efficacy.To manifest this issue, efficient, and low-cost methods are required to recognize crop diseases accurately.Nowadays, AI-based ML and DL techniques play an essential role in classifying crop leaf attacks by learning the features accurately.Some existing techniques depict the use of ML techniques for the classification of crop leaf attacks.However, ML techniques intensify the error as well as the training time while processing with larger leaf images.Hence, DL techniques are presented and investigated for detecting crop leaf diseases.It is proven that DL techniques have greater capability in processing with larger databases and require minimal features for the disease recognition process.The contributions are: • To present the DL technique for classifying the severity level of leaf diseases in pepper and corn crops.
• To introduce the various stages like dataset collection, noise removal, segmentation, feature extraction, classification and severity prediction.
• To introduce optimal U-net for segmenting and identifying the lesions.
• To introduce hybrid GJ-RD (Golden jackal with Red deer optimizer) and it is the integration of GJA (Golden jackal algorithm) and RDA (Red deer algorithm) for precise detection.
• The rest portion is sorted as section 2 presents the literature works based on the pepper and corn leaves diseases; Framework of the proposed pepper and corn leaves diseases is presented in section 3; section 4 defines the results and the work is concluded in section 5.

Related works
Literature works based on the pepper and corn leaves diseases using different approaches are discussed here: Haque et al [20] defined the leaf disease classification for the pepper using the transfer learning (TL) technique.In this study, the Xception, ResNet152-V2, MobileNet-V2, VGG19, and NasNet Mobile were introduced to classify the pepper leaf diseases accurately.This study classifies multiple classes like Crown Gall, Bacterial Spot, Leaf Spot, Cercospora, Anthracnose, Leaf Curl, Pepper Huasteco, Golden Mosaic, and Gray Leaf Spot.For the simulation process, a GitHub dataset consisting of a total of 386 images was collected and processed in this study.However, this study lacks effective noise-filtering techniques hence the classification error increases.
Akhalifi et al [18] established the fine-tuned TL technique for classifying pepper leaf disease accurately.In this study, three major stages were involved namely pre-processing (resizing, augmentation, and data separation), and classification.Finally, the ResNetV250, triple Fully Connected (FC) layers, MobileNet, and VGG16 were introduced to classify the pepper leaf diseases efficiently.A publicly collected Plant-village database that consists of 997 bacterial spot images and 1478 healthy images was utilized for the experimentation process.However, this method faces high data redundancy problems during the training process.
Andersson et al [21,22] put forth a DL model for the classification of pepper leaf diseases efficiently.Two major stages such as pre-processing and classification were performed.In the pre-processing stage, image resizing, rescaling, and flipping were performed.Then, the DL-based CNN model was introduced for the identification of pepper leaf diseases.In total, 2080 bacterial spot images and 1881 healthy images were utilized for the simulation process.However, this method increases the error while training with larger images.
Mathew et al [23] introduced the DL model for detecting bacterial disease in pepper plants accurately.In this study, various DL techniques like AlexNet, VGG-16, and VGG-19-based techniques were introduced to detect pepper leaf diseases.The pepper leaf images were collected and processed via the Plant-Village database of diseased and healthy images.However, this method lacks optimization strategies leading to high gradient insufficiency problems.
Kundu et al [24] determined the DL technique for detecting, predicting, and loss estimation for corn crops.In this study, k-means clustering (KMC) was introduced to segment the diseased region accurately.Then, a MaizeNet-based customized DL model was introduced to classify the severity level and estimate the crop loss effectively.For experimentation, the dataset was collected from the Indian Council of Agricultural Research-All India Coordinated Research Project center (Mysore) that consisted of 2996 images of different classes.However, this method considered less data for the training process and is likely to cause under-fitting issues.
Sibiya, and Mbuyu [25] defined the DL technique for analyzing the severity level of maize crop in maize crop.Initially, a fuzzy-based threshold segmentation technique was utilized to extract the diseased leaf region based on fuzzy rules.Then, the VGG-16-based DL technique was introduced to classify the severity level of common rust disease in maize crops.A publicly collected Plant-village database was utilized that consists of the severity level of common rust diseases in early, mid, and later stages.However, this method was highly cost-effective and requires a high number of features to learn the severity ratings accurately.

VGG-16 Feasible model and enhanced robustness
This method was highly cost-effective and requires a high number of features to learn the severity ratings accurately Phan et al [26] SLIC Overcome over-fitting issues This method causes high data redundancy problems due to the lack of an effective fine-tuning process Divyanth et al [27] UNet-DeepLabV3+− Determined the severity of multiple diseases The generalization capability was poor for this method and increased the error while training with larger datasets Cui et al [28] CBAM Reduce the dimensionality features in an efficient way CBAM can consume significant memory, especially when dealing with high-resolution images or large batch sizes, which may limit its applicability in resource-constrained environments Yu et al [29] KMC Enhanced the classification performance The generalization capability Phan et al [26] put forth clustering-based segmentation and DL-based techniques for analyzing the diseased regions in corn leaves.In this study, two stages of segmentation and classification were performed to identify the corn leaf diseases accurately.Initially, a Simple Linear Iterative Clustering (SLIC) based segmentation approach was introduced to cluster the diseased region from the corn leaf.Then, various transfer learning techniques were utilized for identifying the corn leaf diseases.However, this method causes high data redundancy problems due to the lack of an effective fine-tuning process.
Divyanth et al [27] introduced the two-stage DL technique for analyzing corn leaf diseases using corn field images.In this study, two stages of segmentation and classification was performed to identify the corn leaf diseases accurately.Initially, the UNet-DeepLabV3+-based combined segmentation and classification approach was introduced to cluster the diseased region and to classify the corn leaf diseases with minimal time complexity.Moreover, Xception with an atrous convolution module was introduced to extract the useful diseased features effectively.The dataset was collected from Purdue University's Agronomy Center for Research and Education (ACRE) consisting of about 1050 images that were utilized for the training process.However, the generalization capability was poor for this method and increased the error while training with larger datasets.Table 1 presents the overview of the literature works.
Cui et al [28] developed CBAM (convolutional block attention model) with autoencoder was utilized for corn leaf disease classification.For reducing the dimsnionality, the DWT (discrete wavelet transform) was presented.CBAM employs attention mechanisms to focus on salient features in both spatial and channel dimensions within feature maps during encoding.In the decoding phase, the Latent space serves as input for reconstruction.
Yu et al [29] presented KMC with dfferernt DL models for corn leaf disease classification.The analysis was performed by varying the values of k and different DL models like VGG16, 19, Inception v3 and ResNet 18 were considered.Average accuracy values achieved by the VGG16, and ResNet 18 were 84.4% and 83.7%.

Research gaps
Mostly, the existing studies failed to utilize the essential pre-processing stage to enhance the image quality resulting in increased training time, and inaccuracy.Moreover, an effective segmentation technique is lacking in existing studies to analyze the severity level of crop leaves.The existing practices show poor accuracy performance and hence, they highly require the support of human interventions.Also, when the number of crop images is enlarged, the training fluctuations in the network model are increased.To solve these problems, an effective and lightweight technique is highly required to achieve accurate classification performance.To our knowledge, the proposed study overcomes all the problems faced by the existing studies and provides outstanding performance effectively.

Proposed methodology
In the field of agriculture, diagnosing plant diseases and estimating their severity are the significant challenges.In this study, we introduce a reliable plant disease diagnosis and severity estimation model.This section explains and examines the suggested methods for categorizing lesions conditions.The following components make up the entire process: pre-processing, the second phase is segmentation, followed by feature extraction, classification and the last step is severity prediction.To overcome the hand-crafted segmentation and feature extraction this work presents an optimal U-Net model for identifying the corn and pepper leaves.Figure 1 presents an overview of the suggested model.

Pre-processing
The occurrence of noise image may affect the classification performance.Hence, it is essential to achieve noise free images.For removing the noise in the input grayscale leaves images, the median filtering is utilized.This filter removes noise as well as preserves the edges.The median filter uses a pixel with the median range of every neighbouring pixels.Then, the value of pixel in specific location is replaced by the median pixel value and it is given as: where ( ) I p q , is the output of the image, u is the neighbourhood that the user has defined, ( ) K L , is the image's location.Hence, by the median filtering approach the noise free images with smooth images are obtained.

Optimal segmentation and classification
The enhanced images are given as input to the optimized U-Net model.The structure of U-Net is given in figure 2 and it has two segments like encoding (up-sampling) and decoding (down-sampling).The encoder and decoder models are integrated to make the optimized U-Net model.The standard CNN doesn't extract the semantic information in an efficient manner.In the optimized U-Net model, decoder part collects the semantic information from the lowest layer of the network.The fine segment structures are provided by the decoder part, which receives data directly from the encoder part using the by-pass connection.
There are five convolutional blocks in the decoding and each block has two convolutional layers having the 3 × 3 filter size.The activation function utilized is ReLU and the feature maps are ranges from 1 to 1024.For the decoding path, max-pooling with 2 × 2 strides is provided at the ending segment of each block.Hence, the feature maps size is reduced to 15 × 15.In the encoding path, each block begins with a deconvolutional layer that has a 3 × 3 filter size and a 2 × 2 stride.In this encoding path, reduces the number of feature maps by two while doubling their size in both directions.Hence, the feature maps size increases to 240 × 240.In each encoding block, 2 convolutional layers minimize the deconvolution feature maps.Then, the zero padding is utilized in both encoding and decoding for keeping the output dimensions for every convolutional layer.Then, to reduce the number of feature maps 1×1 convolutional layer is utilized.

GJA (Golden jackal algorithm)
The Golden Jackal Algorithm is based on the hunting behaviour of golden jackals, which are known for their cooperative hunting techniques and adaptability to changing environments.The GJA effectively optimises solutions in search spaces by mimicking jackals' collaboration and adaptive nature.It takes a population-based approach, in which each jackal (solution) interacts with its neighbours to share information and adapt its position in response to local and global influences.This collaborative behaviour allows the algorithm to quickly converge on optimal or near-optimal solutions.
GJA model portrays the hunting characteristics of the GJ (Golden jackal) and it is the terrestrial predator.The hunting characteristics of GJ are mathematically developed for building an optimization.The GJ is the metaheuristic optimizer and the initial candidate's solution ( ) Y 0 is identified by the search space.
where is the random number, and are upper and lower limits.The jackal pair is selected by the following expression where Y k l , is the l th element of k th prey, n and D are the total preys and number of parameters.The fitness of prey ( ) Fit prey is determined by the optimization and it is expressed as: Exploration stage: GJ can identify and capture the prey; the hunting process of GJ is expressed as: ´-ẃhere t is the present iteration and ( ) prey t position vector of prey, ( ) Y t M and ( ) Y t FM are the male and female GJ location.Prey's escaping energy E n is given as: = ẃhere E 0 and E 1 are the initial and final energies.Then, the random number r is computed by: ( ) ( ) r L e v y y 0.5 8 = ẃhere ( ) Levy y is the levy's flight.BY using the (5) and (6), the position of GJ is updated by:

+ = +
where ( ) Y t 1 1 + is the prey's location updated by the male and female GJ.Exploitation stage: When prey is harassed by jackals, its ability to escape is minimized and the jackal pairs surround the prey.After the encircling process, the GJ jump over the prey, and consume it.This hunting strategy along with the male and female GJ is given as: Switch from exploration to exploitation: In GJA, switching from exploration to exploitation E n is used.The prey becomes weak when E 0 drops from 0 to −1, but it gets stronger when E 0 increase from 0 to 1.When | | E 1, n > male and female GJ search for prey in various locations for exploring prey.When | | E 1, n < male and female GJ attack and exploit the prey.

RDA (Red deer algorithm)
The Red Deer Algorithm was inspired after the hierarchical social structure and mating behaviour of red deer.Red deer exhibit complex social interactions and dominance hierarchies, which have been emulated in the RDA to develop a robust optimisation algorithm.In the RDA, each deer (solution) has a hierarchical position within the population, and interactions are governed by dominance and competition.This hierarchical structure allows balanced exploration of the search space, ensuring both the exploration of new regions and the exploitation of promising areas.
RDA model portrays characteristics of the RD (red deer) and it initiates by the initial population.The best RD on the population is considered as MRD (male RD) and the remaining RDs are considered as hinds.There are two types of RDs like leaders and stages are considered.The mathematical computation of the RDs is formulated in the following section.
Roaring of male RD: In roaring stage, MRD tries to improve the elegance and obtain better optimal solution.Neighbor of the MRD and the objective solution of neighboring are higher than the MRG, then the previous solution is utilized.For updating the males' position, the below expression is utilized.
) ( ) male male c ul ll c ll when c male c ul ll c ll when c 0.5 0.5 12 where c c c , , 1 2 3 are the random numbers, male o and male n are the old and new MRDs, ul and ll are the upper and lower limits.
Selection of best MRD as male commanders: There are two kinds of RD; they are leaders and stags.The total number of leader males G Lead is expressed as: Lead male a = ẃhere a is the initial value.Then, the total number of stags G stag is expressed as: Conflict among stags and male leaders: When the leaders and stags conflict with each other, the new solution is obtained.Then, the solution of leader is replaced with the new one and it is given as: where N 1 and N 2 are the two new solutions.u 1 and u 2 are the uniform distribution terms.Formation of harems: The strength of the male leaders in harems determines how many hinds are there.For creating the harems, the hinds between leaders are represented as: where v m is the strength of m th leader and V m is the optimized value.For computing the strength of the male leaders, the following expression is utilized.
The number of harem's hinds harem m is expressed as: ´= ẃhere M hind is the number of hind.Mating: The male leader RD mates with other leader and harem m mate with other leader is represented as: b is the initial value and the process of mating is given as: a is the random number.Mating harem commander with other harem: The number of hinds in the l th harem and its mating process is given as: stag with the nearby hind: In this stage, every stage mate with the nearby hind.For finding the hind, the distance Dis j among every hind and the stage in N-dimension space is computed by:

Proposed hybrid GJ-RD algorithm
This work presented hybrid GJ-RD algorithm to segment the lesions in corn and paddy leaves.To enhance the segment the accuracy, the hyper-parameters of the UNet (learning rate, drop out and batch size) is optimized by the proposed hybrid GJ-RD.The basic model of the GJA is increased by improving the position update of population.This hybridization ensures the GJA greater flexibility in identifying the population, variability and reaches the optimal values.

Application in classification of pepper and corn leaves
In the context of classifying and identifying the severity of diseases in pepper and corn leaves using a hybrid optimization-based U-Net model, GJA and RDA can be employed to optimize the model parameters and enhance its performance.The hybrid optimization approach combines the strengths of both algorithms to achieve better convergence and solution quality.GJA can be utilized to explore the feature space efficiently, identifying relevant features that distinguish between healthy and diseased leaves.Its collaborative nature can help in fine-tuning the parameters of the U-Net model to improve classification accuracy.
RDA can be employed to optimize the hyperparameters of the U-Net model, such as learning rates, batch sizes, and network architecture.Its hierarchical structure ensures a balanced exploration and exploitation strategy, leading to robust and stable model performance.
By integrating GJA and RDA into the optimization process of the hybrid U-Net model, the classification and severity identification of pepper and corn leaves can be enhanced significantly.These bio-inspired algorithms offer innovative solutions to complex optimization problems, providing a promising avenue for improving agricultural image analysis and disease diagnosis.The robustness of every solution is evaluated by computing the fitness and selects the best solution.

( ) ( ) Fitness Max Accuracy 24 =
The proposed hybrid GJ-RD algorithm starts by initialization of GJA and RDA parameters.Then, the proposed algorithm undergoes the updation of present population by either GJA or RDA.When the stopping criteria are met, then the further stage is to moves to the better solution.Figure 3 shows the flowchart of the proposed hybrid GJ-RD algorithm.

Severity estimation
This section shows how the lesions severity present in the pepper and the corn leaves is identified.It is critical for farmers to know whether pests or diseases are present.To avoid losing crop yields, it is crucial to understand the disease's severity and take appropriate action.The following expression is used for identifying the severity.
Area d is the diseased area in the corn and pepper leaves and Area t is the overall leaves area of corn and pepper.

Results analysis
Different leaf images from pepper and corn crops are used to train the DL-based plant disease segmentation, and classification process.The device has an Intel Xenon processor running 64-bit Windows 10, 64 GB of RAM, and Python programming tool is utilized.Table 2 presents the parameter setting of the proposed pepper and corn leaves disease classification.

Ablation study
Following section shows the ablation study of the various approaches like UNet, UNet-RDA, UNet-GJA and the proposed disease classification model.Then, the convergence analysis, accuracy-loss curve, ROC and confusion matrix are presented.Finally, the severity is determined for the proposed disease classification model.Figure 7 depicts the ROC-AUC (area under the curve) comparison of the approaches.The AUC values achieved by the UNet, UNet-GJA, UNet-RDA and the proposed UNet-hybrid GJ-RD algorithm are 0.92, 0.94, 0.97 and 0.99 on the PlantVillage dataset.
Figure 8 presents the converge analysis and the performance is taken for the optimization approaches like RDA, GJA and the proposed GJ-RD.for all iterations, the fitness values are constant throughout the fitness values.However, the fitness of the proposed GJ-RD is very less than the other two optimization approaches.
Figure 9 depicts the accuracy-loss curve of the proposed UNet-hybrid GJ-RD algorithm.The performance is evaluated by varying the epoch values of 100 with respect to the training and testing accuracies.It is proved that      GJA, UNet-RDA and the proposed UNet-hybrid GJ-RD algorithm are 91.6%,93.87, 94% and 96.8% respectively.
Figure 11 defines the recall comparison of the various DL models by varying the training percentage from 60 to 100.It is proved from the graph that the when the training percentage is increased, the recall performance of all approaches are also increasing.When the training percentage is 70 and 80, the recall values achieved by the proposed UNet-hybrid GJ-RD algorithm are 97.06% and 97.08%.When the training percentage is 90, the recall value achieved by the proposed UNet-hybrid GJ-RD algorithm 98% and it is 1.01%, 2.1% and 2.2% better than the DL models like UNet, UNet-GJA and UNet-RDA.
Figure 12 shows the precision comparison of the various DL models by varying the training percentage from 60 to 100.The precision values achieved by the proposed UNet-hybrid GJ-RD algorithm are 97.8% and 99.18%.The proposed disease classification achieved due to the hybridization of the GJA and RDA.Further, the conventional models achieved poor results due to slow convergence and overfitting issues.
Figure 13 shows the F score comparison of the various DL models by varying the training percentage from 60 to 100.It is observed that the F score values achieved by the proposed UNet-hybrid GJ-RD algorithm are 95.9%,96.7%, 97.1%, 98% and 99.8% on the PlantVillage dataset.
Figure 14 shows the specificity comparison of the various DL models by varying the training percentage from 60 to 100.It is observed that the specificity values achieved by the proposed UNet-hybrid GJ-RD algorithm are 99.6% on the PlantVillage dataset.
Thus, it is proved that the proposed UNet-hybrid GJ-RD algorithm is suitable for pepper and corn leaves diseases classification.
Table 3 presents the comparative analysis of the various approaches with respect to the pepper and corn leaves diseases.It is observed from all analyses that the proposed UNet-hybrid GJ-RD outperformed the conventional models.This betterment is due to automated segmentation, feature extraction and classification by the UNet-hybrid GJ-RD.this hybrid optimization enhance the convergence speed of the algorithm.

Conclusion
This work presented an automated optimization-based DL approach for corn and pepper leaves disease segmentation and classification.To overcome the hand-crafted feature extraction and manual segmentation, the DL model optimal U-Net was presented.For enhancing the performance, the algorithm hybrid GJ-RD was utilized.The experimental analysis proved that the proposed optimal U-Net model achieved better accuracy, precision and recall values of 99.2%, 99.1% and 99.03% on the PlantVillage dataset.Finally, the severity estimation was determined to estimate the severity of lesions.Linear regression plot of various classes like CGLS, CR, CNLB, CH, PBS and PB were also presented.The proposed lesions segmentation and severity estimation model assists the farmers and increase the crop productivity.In the future, the suggested system could be a good fit for a portable device that diagnoses plant diseases in real time.Moreover, future research will concentrate on real-time segmentation and classification with the aid of more effective methods like ensemble learning and DL models.Moreover, different optimization algorithms with fine-tuned DL models will be applied for classifying diseases.Our suggested approach can be used by researchers in different fields to manage processes in various application domains.

Figure 1 .
Figure 1.Overview of the suggested model.

=
Choose the next generation: Two mechanisms are utilized for selecting the next generation.In the first mechanism, every MRD, stag and leaders are involved.The second mechanism represents the remaining population.

Figure 3 .
Figure 3. Flowchart of the proposed hybrid GJ-RD algorithm.

Figure 4 .
Figure 4. Sample images of the dataset.

4. 1 .
Dataset analysesThe Plant Village dataset[30], which includes both healthy and unhealthy (bacterial spot) cases, is used for experimentation in this work.Images of fourteen plants are included in this dataset, and the study examined pepper and corn leaves diseases.The data in this dataset has a size of 37.3 MB, and the dataset as a whole is approximately 857 MB.In the corn dataset, there are 282 CGLS (cercospora leaf spot gray leaf spot) images, 538 CR (common rust) images, 342 CNLB (corn northern leaf blight) images and 430 CH (corn healthy) images.Further, in the pepper dataset there are 301 PBS (pepper bacterial spot) images and 446 PB (pepper bell) images.Figure4depicts the sample images of the dataset.

F1 score:
It is the average mean of P and R. It is represented as: It is the ratio of leaves that are wrongly recognized by the classifier and it is represented as: Qualitative analysis Figure5shows the qualitative representation of (a) input image, (b) Resized image, (c) Pre-processed image and (d) Segmented

Figure 6
depicts the confusion matrix of the different algorithms.Here, 0 indicates the CGLS, 1 indicates the CR, 2 indicates the CNLB, 3 indicates the CH, 4 indicates the PBS, and 5 indicates the PB.In this matrix, 57 samples are classified as 0, 119 samples are classified as 1, 62 samples are classified as 2, 79 samples are classified as 3, 56 samples are classified as 4 and 87 samples are classified as 5 for the the proposed UNet-hybrid GJ-RD algorithm.

Figure 9 .
Figure 9. Accuracy-loss curve of the proposed UNet-hybrid GJ-RD algorithm.

Figure 10
defines the accuracy comparison of the various DL models.The methods like GJ-RD are compared with the proposed UNet-hybrid GJ-RD algorithm.The performance is analyzed by varying the training percentage from 60 to 100.When the training percentage is 60, the accuracy values achieved by the UNet, UNet-

Table 1 .
Overview of the literature works.