3.1 Dataset Description

The research utilized cotton leaf images from Kaggle (Kaggle.com 2025), which included 2637 images logically classified into aphid, army worm, bacterial blight, healthy leaf, powdery mildew, and target spot categories. Different light levels and natural environmental factors were utilized to collect 2637 images of contaminated and uninfected cotton leaves, which created suitable training conditions for the model. The images are of high-definition quality with added labels to enable supervised training. The information is properly organized through distinct folders that represent each disease type. Figure 2 shows the dataset samples of all classes. Aphids that harm cotton yield and plant health inflict damage on the leaves, which presents four affected samples. Small pests known as aphids construct clusters under leaf surfaces where they drain sap fluid and cause foliage discoloration, leaf distortion, and reduced plant growth. The dataset shows high-quality pictures that demonstrate the diverse effects of bug infestations while revealing different environmental aspects. The divergent elements in the dataset enable training artificial intelligence models to correctly identify and categorize leaves with aphid infestations. Multiple perspective data in the dataset reinforce machine learning applications while permitting exact pest detection, which supports automatic pest management tools for sustainable cotton farming. A total of four cotton leaf samples demonstrated army worm attack effects, which extensively degraded the foliage of cotton plants. The feeding pattern of army worms comprises several destructive traits, including fast spread and heavy consumption, which results in leaf damage patterns that become skeletonized with reduced photosynthetic function. The diverse infestation levels and environmental specifications in the data enable DL algorithms to detect a wide range of army worm damage manifestations. New high-quality images help machines learn better, which has led to the development of automatic early pest warning systems and targeted agricultural techniques for cotton fields (Ramos et al. 2022).

Bacterial blight ailments in cotton leaves display visible symptoms of water-soaked lesions and necrotic spots, as well as vein browning. DL models benefit from various stages of bacterial infection visible in the collected samples for advanced training capabilities. The occurrence of this disease is triggered by Xanthomonas citri pv. malvacearum spreads through rain splashes and contaminated seeds, which makes early disease detection essential. The detailed images in the dataset enable accurate identification of distinct features, which allows precise classification processes. Using these images, the diagnostic ability of machine learning models can be enhanced to detect cotton diseases immediately and enable effective disease management systems. The healthy leaf samples were used as part of the control data for disease classification model evaluation. The healthy leaf images display clear leaf outlines together with a uniform color scheme with no indications of disease or pest infestation. The images were obtained under different environmental conditions and lighting scenarios, which resulted in better model accuracy and generalizability. Widely used for DL framework training, the healthy leaf dataset helps distinguish between infected and uninfected cotton leaves, thus reducing false alerts during disease detection model operations. Using high-quality, healthy leaf images during AI-based agricultural system training produces accurate disease classification capabilities for cotton cultivation (Anwar et al. 2022). Powdery mildew is a fungal disease that results in white powdery mildew growth on cotton leaves, which damages plant photosynthesis and affects the health of the plant system. The infection severity scales from mild spots through multiple stages of infection to complete leaf coverage and appears across the high-resolution images present in the database. The fungus spreads quickly when humidity is present; therefore, detecting its presence early becomes essential. Because of this diverse sample collection, DL models have become better at detecting fungal infections while discriminating them from other leaf diseases. This leads to improved disease prediction and prompt farmer intervention. The appearance of the target spot in the cotton leaves revealed circular, dark brown spots surrounded by yellow areas. Fungal infection severely harms both cotton production levels and product quality. The dataset contains different infection grades, which helps DL models detect early-stage infections and separate them from symptoms associated with other pathogens. These detailed images enable researchers to perform feature analysis while dividing images and classifying objects, which leads to superior disease diagnosis and better control methods for sustainable cotton production (Xavier et al. 2019).

3.2 Data Preprocessing and Augmentation

The performance, generalizability, and consistency of the proposed model of DL used to classify cotton leaf diseases were improved by a clearly defined preprocessing pipeline that was intended to streamline the process of disease classification. The original dataset was obtained through Kaggle, with high-resolution pictures being organized into six classes. In order to fix the size of input data and palliate the calculation load, every picture was resized to 224 × 224 pixels. This dimension standardization allows compatibility with the input requirements of EfficientNetB3 and InceptionResNetV2, the two backbone architectures upon which the feature extraction will happen. Min-max scaling was used as normalization: all values of pixels were transformed to the interval [0, 1]. This step improves upon the convergence of the model and numerical stability when training the model, especially in deep networks where unnormalized data may give gradient instabilities. Several data augmentation methods were used to enhance the generalization of the trained model and avoid overfitting. Among them, there were random rotations with a maximum angle of 25°, horizontal flipping, shifting the width and height by 20%, and zooming to a maximum of 30%. These augmentations have been done on only the training data in a way that kept the labels of the classes the same, but gave each class varied representations. This plan is more representative of field variability since the model is resistant to changes in lighting, orientation, and position of leaves, as in the real-world situation. Notably, this increase enables the model to learn invariant representations, which richly decrease its dependency on fixed positioning cues and, therefore, increase its flexibility under unseen circumstances. In unison, these preprocessing methods established the basis of a scalable, high-performing, and real-time deployable classification model that could be used in precision agriculture. The preprocessing pipeline allowed solving the dataset imbalance and variability, which is why it brought a substantial improvement in accuracy, recall, and robustness presented by the final model (Manavalan 2022).

The original images obtained from the Kaggle dataset were of high resolution, with dimensions of 256 × 256 pixels, which were uniformly resized to 224 × 224 pixels to ensure compatibility with the input layers of EfficientNetB3 and InceptionResNetV2. The choice of a 25° rotation range was made empirically as a moderate augmentation level that preserves key leaf features while still introducing meaningful variance in orientation. This range was found to simulate realistic leaf rotations observed in natural environments without causing distortion or excessive deformation of disease-specific patterns. Importantly, random rotations were applied in both clockwise and counterclockwise directions to enrich directional variability. To accommodate the values of the pixels that accompanied the transformation (e.g., rotation, shifting, resizing), we resorted to constant black padding, which guaranteed minimal obscuring of the image information and constant dimensions of this entire image. The preprocessing steps that were selected carefully contributed to the model's generalization and correspond closely to the variability experienced in cotton field imagery.

The data were further stratified with an 80:20 ratio and partitioned into training and validation. The stratification maintained the distribution of classes in the two sets and promoted an equal demonstration of every one of the six-category diseases in the training and assessment stages. The separation was done via the train_test_split() algorithm of Scikit-learn, with stratify = y_resampled that ensures that the proportions of classes are preserved. Moreover, the test set was externally retrieved through an existing directory in the Kaggle collection and was not subject to resampling or augmentation techniques. This plan avoided such leakage of data and gave an objective assessment of model performance.

A hold-out validation method was embraced to determine the model's generalization and prevent overfitting. In particular, due to the class imbalance, SMOTE was used to resample the training data into training and validation sets as part of the 80:20 stratified split. It enabled us to test the model on a representative subset and retain the class distribution. It also used early stopping due to validation loss during training, so as to stop overfitting and achieve fast convergence. This validation strategy allowed all the experiments to yield consistent results and allowed reliable hyperparameter tuning in the GA optimization process.

3.3 Class Imbalance Handling With SMOTE

Class imbalance is also a long-standing problem in agricultural disease detection, where some types of the disease are not well represented in the training data. Such skew leads to DL models being biased toward the majority classes in the training process, frequently at the cost of accurately classifying minority class samples. The training set in the cotton leaf disease used in this research was highly imbalanced, with the number of samples representing certain classes, like bacterial blight or army worm, being considerably smaller compared to others, such as aphids or healthy leaves. This class imbalance lowered the performance of the classification procedure, especially in the rare types of diseases, since the model was constrained to associate representative patterns of all the classes (Ramanjot et al. 2023). As a solution to this problem, we used the Synthetic Minority Oversampling Technique (SMOTE). SMOTE is a widely accepted oversampling method that generates synthetic data points for minority classes by interpolating between existing samples in the feature space. Unlike naive oversampling, which duplicates existing data, SMOTE creates new, unique samples, improving generalization and reducing overfitting risks. It was applied exclusively to the training data after feature extraction to preserve test data integrity and avoid data leakage. Visual evidence of class imbalance and its correction is provided through t-SNE plots and bar charts. As shown in Figure 3, before SMOTE application, the t-SNE visualization reveals sparse data point densities for underrepresented classes such as bacterial blight (Class 3) and army worm (Class 1). These class clusters appear disconnected and limited in sample representation, making it difficult for the model to extract meaningful features. In contrast, Figure 4 demonstrates how SMOTE produces a more uniform class distribution. Synthetic samples generated for minority classes result in denser and more consistent t-SNE clusters, enabling balanced learning across categories (Chawla et al. 2002).

Additionally, Figure 5 presents class distribution comparisons before and after SMOTE. The original dataset contained uneven class frequencies, with as few as 27 images in some classes. Post-SMOTE, each class was equalized to 39 samples by synthesizing: 12 new images for bacterial blight, 8 each for army worm and healthy leaves, and 3 each for powdery mildew and target spot. The SMOTE technique effectively removed class-based bias, improving the model's ability to learn distinct patterns for every disease category. This contributed to higher classification accuracy and better per-class precision and recall, especially for minority classes that were previously misclassified. While alternative strategies like class-weighted loss or focal loss are effective during training, they adjust loss contribution rather than data distribution. In this study, SMOTE was combined with focal loss to both equalize the training data and focus the model's attention on difficult-to-classify samples. This dual approach strengthened minority class learning while maintaining robustness for well-represented classes. The application of SMOTE in the proposed pipeline not only improved class distribution but also significantly enhanced the model's generalization, fairness, and diagnostic performance across all disease types. It plays a vital role in ensuring real-world applicability where class imbalance is often unavoidable in agricultural datasets (Bhagat and Kumar 2022).

3.4 Proposed Model Architecture

Figure 6 shows the proposed model architecture for cotton leaf disease classification (Kursun and Koklu 2024), which combines advanced feature extraction methods with optimization approaches and classification algorithms. Two pretrained CNNs process the input dataset through InceptionResNetV2 and EfficientNetB3. The networks make use of global average pooling (GAP) to obtain significantly high-dimensional feature representations from input images. After extraction, these features are placed in a unified space to increase the discrimination ability for different cotton leaf disease types. A GA serves the purpose of hyperparameter optimization to maximize model performance. Three essential model parameters are optimized from the GA framework to help the model achieve its best accuracy while improving the generalization outcomes. The external layers include batch normalization together with dropout and dense layers, which make the model more resistant and minimize overfitting effects. The Softmax classifier serves as the final component by sorting leaf conditions into six major categories, where both uninfected and infected leaves receive proper classification. The model design enables reliable and easy-to-read performance, which makes it suitable for practical precision agriculture deployments. The automated disease detection system, powered by accurate classifications offered by this model, enables farmers and agronomists to make better decisions, thus promoting sustainable agricultural practices.

3.4.2 Hyperparameter Optimization Using Genetic Algorithm

DL model performance and the ability to generalize are centrally dependent on hyperparameter tuning. This paper resorts to a GA (Latif et al. 2021), which is a population-based metaheuristic, and is based on the concepts of natural evolution, to automate this process and to realize the best possible configurations. Optimization of the important hyperparameters of the fully connected classifier, that is, learning rate ($$$ \alpha $$$), the number of neurons ($$$ \mathcal{u} $$$) in the dense layer, and the dropout rate ($$$ \mathcal{d} $$$), is performed using the GA. The method allows the effective search of high-dimensional and nonlinear hyperparameter space, without the use of gradient information. GA compares well in attributes when compared to other methods, for example, Bayesian optimization or Particle Swarm Optimization (PSO) in that it is more robust in processing discrete and continuous parameters concurrently and is less susceptible to local optima. Its random appearance leading to larger search power is also caused by its stochastic nature and capability to retain diverse population throughout a series of generations.

This is started by randomly initializing a population of hyperparameter sets:

$$$$ {\theta}_i=\left\{{\alpha}_i,{\mathcal{u}}_i,{\mathcal{d}}_i\right\},\kern1.75em \mathcal{i}\in \left\{1,2,\dots, N\right\} $$$$(6)

Each individual $$$ {\theta}_i $$$ in the population is evaluated by training the model for a limited number of epochs and computing validation accuracy as a fitness score:

$$$$ A\left({\theta}_i\right)=\frac{\sum_{j=1}^N\mathbbm{I}\left({\hat{y}}_j={y}_j\right)}{N} $$$$(7)

Here, $$$ {\hat{y}}_j $$$ and $$$ {y}_j $$$ denote the predicted and actual class labels, and $$$ \mathbbm{I}\ (.) $$$ is the indicator function. The top-performing half of the population is selected, and crossover is applied to mix traits from high-fitness individuals. Mutation is then introduced by adding Gaussian noise to promote diversity:

$$$$ {\theta}^{\prime }=\theta +\in, \kern1.75em \in \sim \mathcal{N}\left(0,{\sigma}^2\right) $$$$(8)

This evolutionary cycle is repeated for $$$ G $$$ generations until convergence is achieved or no further accuracy gain is observed. The best solution is then used to train the final model on the full training dataset.

Algorithm 1 outlines the entire GA optimization procedure, including initialization, fitness evaluation, selection, crossover, mutation, and termination. The hyperparameters discovered by the GA contributed significantly to performance improvements, reducing overfitting, and maximizing the classification accuracy, precision, and recall on both validation and test datasets.

ALGORITHM 1. Genetic optimization-based deep learning model for cotton leaf disease classification.

Input:

• Preprocessed dataset $$$ {D}_{\mathrm{train}},{D}_{\mathrm{val}},{D}_{\mathrm{test}} $$$
• Number of generations $$$ G $$$
• Population size $$$ P $$$
• Learning rate range $$$ \alpha \in \left[0.00005,0.003\right] $$$
• Number of units in dense layers $$$ u\in \left[128,512\right] $$$
• Dropout rate $$$ d\in \left[0.2,0.5\right] $$$

Output:

• Optimized hyperparameters for DL model $$$ {\theta}^{\ast }={\alpha}^{\ast },{u}^{\ast },{d}^{\ast } $$$
• Trained DL model for cotton leaf disease classification

Step 1: Genetic Algorithm Initialization

• Initialize a population $$$ P $$$ of size $$$ N $$$, where each individual represents a set of hyperparameters $$$ {\theta}_i $$$ = {$$$ {\alpha}_i,\kern0.5em {u}_i,\kern0.5em {d}_i $$$}.
• Randomly assign values to each hyperparameter within their respective ranges.
• Initialize EfficientNetB3 and InceptionResNetV2 as feature extractors with pretrained ImageNet weights.
• Remove top layers from both models and use GAP for feature extraction.
• Concatenate extracted feature vectors from both models.
• Add Dense layers, Batch Normalization, and Dropout for regularization.
• Initialize a Softmax output layer to classify into six classes: Aphids, Army Worm, Bacterial Blight, Healthy Leaf, Powdery Mildew, and Target Spot.

Step 2: Model Training and Fitness Evaluation

For each generation $$$ g $$$ from 1 to $$$ G $$$:

For each individual $$$ {\theta}_i $$$ in the population $$$ P $$$:

• Train the DL model using a training dataset $$$ {D}_{\mathrm{train}} $$$ with hyperparameters $$$ {\theta}_i $$$.
• Compute validation accuracy $$$ A\left({\theta}_i\right) $$$ as the fitness score:
  $$$$ A\left({\theta}_i\right)=\frac{\sum_{j=1}^N\mathbbm{l}\left({\hat{y}}_j={y}_j\right)}{N} $$$$

where $$$ N $$$ is the number of validation samples, $$$ {y}_j $$$ is the true label, and $$$ {\hat{y}}_j $$$ is the predicted label.

Step 3: Selection of Best Individuals

• Rank individuals based on their fitness scores $$$ A\left({\theta}_i\right) $$$.
• Select the top 50% of individuals with the highest validation accuracy:
  $$$$ S=\left\{{\theta}_k|A\left({\theta}_k\right)\ge {A}_{\mathrm{median}}\right\} $$$$

Step 4: Crossover (Generating New Individuals)

• Generate offspring by averaging the hyperparameters of two selected parents:
  $$$$ {\theta}_{\mathrm{child}}=\frac{\theta_{\mathrm{parent}1}+{\theta}_{\mathrm{parent}2}}{2} $$$$
• Ensure that the newly generated hyperparameter values remain within predefined ranges.

Step 5: Mutation (Exploration of New Hyperparameters)

• Introduce random variations in offspring hyperparameters to prevent premature convergence:
• The mutation rate determines how often hyperparameter values are modified.
  $$$$ {\theta}^{\prime }=\theta +\in, \kern1.75em \in \sim \mathcal{N}\left(0,{\sigma}^2\right) $$$$

Step 6: Evolutionary Process

• Replace the old population with the new generation.
• Repeat Selection, Crossover, and Mutation until reaching the final generation $$$ G $$$.

Step 7: Final Model Selection and Deployment

• Select the best hyperparameter set:
  $$$$ {\theta}^{\ast }=\arg \underset{\theta }{\max }A\left(\theta \right) $$$$
• Retrain the final model using $$$ {\theta}^{\ast } $$$ on the full dataset $$$ {D}_{\mathrm{train}} $$$.
• Evaluate the trained model on the test dataset $$$ {D}_{\mathrm{test}} $$$ using:

Accuracy, Precision, Recall, and F1-Score

• Save the trained model for deployment in real-world agricultural applications.

Figure 7 depicts the operational process of the genetic algorithm, which shows how the system continuously develops better hyperparameters to optimize DL model outcomes. The evolutionary methodology effectively optimizes repetitive parameters, which results in more accurate classifications at faster convergence speeds and greater generalization abilities for cotton leaf disease identification.

In our implementation, the GA was configured with a population size of 20, a mutation rate implicitly applied through random resampling in each generation, and a total of 15 generations to evolve the optimal hyperparameter configuration. Each individual in the population encoded three critical parameters for the classifier: learning rate (sampled from 0.00005 to 0.003), number of units in the first dense layer (ranging from 128 to 512), and dropout rate (ranging from 0.2 to 0.5). These ranges were chosen based on standard DL practices for medical/agricultural datasets and refined through empirical tuning. Compared with alternatives like Bayesian Optimization or Grid Search, GA offers superior flexibility in handling both discrete (units) and continuous (learning rate, dropout) hyperparameters simultaneously. Additionally, GA is less prone to local optima due to its stochastic nature and maintains population diversity, which is particularly beneficial in complex, nonlinear optimization spaces like DL model training. This makes GA a robust and scalable choice for metaheuristic tuning in real-world classification problems, including agricultural diagnostics.

3.5 Explainable AI in Cotton Leaf Disease Classification

The complex nature of CNNs used in classification tasks, alongside their vast number of parameters, results in their classification as black-box models. XAI techniques generate model interpretations through explanations that reveal how predictive decisions are made for these models. The two preferred model-agnostic XAI methods currently available include local interpretable model-agnostic explanations (LIMEs) and Shapley additive explanations (SHAPs). The automated explanation feature of these methods demonstrates important model-influencing factors while simultaneously improving human confidence in artificial intelligence systems that classify cotton leaf disease (Sagar et al. 2024).

The interpretability of the proposed model is strengthened through the integration of two widely adopted explainable AI (XAI) techniques: Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP). These tools provide both local (instance-level) and global (feature-level) insights into the decision-making process of the model, enhancing transparency and facilitating trust among domain experts. LIME was used to generate visual explanations for individual predictions by perturbing the input images and observing the corresponding changes in prediction outcomes. Class-wise LIME visualizations were created for representative samples from each disease category (Figure 12), highlighting the specific leaf regions that contributed positively or negatively to each classification. For instance, in bacterial blight, LIME highlighted the water-soaked lesion margins, while for aphids, it focused on discolored areas near the leaf veins. These localized visual cues align closely with agronomic diagnosis criteria, thereby validating the model's attention patterns and reinforcing its credibility for field deployment. SHAP, on the other hand, provided a global understanding of feature importance by attributing predictive influence scores to each feature across the dataset. SHAP summary plots were generated for all six disease classes (Figure 13), and SHAP summary plots of each class are shown in Figure 14, revealing consistent patterns in how features such as lesion shape, color distribution, and texture influenced model decisions. For example, features corresponding to edge sharpness and localized contrast were consistently influential in differentiating target spots from bacterial blight, where the model had previously shown some misclassification. These plots serve as valuable diagnostic guides for agronomists by identifying not only the regions used by the model but also the relative importance of visual patterns across categories. In practice, these interpretability tools empower agronomists and plant pathologists to verify AI-generated diagnoses, identify any potential overfitting (e.g., if the model relies on irrelevant background features), and explore novel disease characteristics not easily visible to the human eye. Furthermore, they can be used to design targeted field surveys and data augmentation strategies. This integration of explainability into the classification framework thus bridges the gap between DL and expert-guided decision-making in precision agriculture.

4.1 Model Evaluation Without the Genetic Algorithm

The DL model's evaluation without the GA utilized several performance metrics, including accuracy, loss, precision, recall, F1 score, and area under the curve (AUC). A training period of 50 epochs allowed the model to reach the final performance evaluation using completely new data. Figure 8 shows the model's comprehensive learning dynamic assessment through six performance indicators, including accuracy, loss, precision, recall, F1 score, and ROC curve assessment. The accuracy curve (Figure 8a) demonstrates a steady improvement throughout the training process. The model begins with 52.32% training accuracy in the initial epoch before achieving 98.09% accuracy during the 50th epoch. The validation accuracy progressively increased to achieve 88.54% precision during the last epoch. The model performs well by learning from training data and applying those types of learning effectively to validation data. The error function reduction pattern throughout the epochs is shown in Figure 8b. During the first epoch, the training loss was detected at 0.043, and it decreased to 0.001 after the 50 epochs alongside the validation loss, which decreased from 0.081 to 0.010. The model effectively reduces the error over time because its loss output remains stable. The precision and recall curves presented in Figure 8c,d show how the model succeeds in producing accurate positive results. The model's precision measure started at 69.39% in the first run and reached 98.55% precision in its final operation. The recall indicator starts at 38.56% and achieves a final percentage of 96.68% during the last epoch. At the beginning of training, the model demonstrated poor precision and recall performance for specific disease classes. The precision and recall values achieved stability throughout training, which demonstrates the development of a proper classification system. The analysis of the F1 score (Figure 8e) demonstrated that the model remained dependable throughout. The model effectively balances the precision and recalls metrics because the F1 score increases steadily, which is the harmonic mean of these two metrics. The F1 score achieved its highest level during the last epoch, thus demonstrating consistent prediction results. The ROC curve (Figure 8f) demonstrates what the model can discriminate when it is used for classification. During 50 epochs, the AUC value began at 0.817 but reached 0.999, indicating exceptional classification effectiveness. The AUC values increase steadily during multiple iterations because the model performs well in terms of true positive identification, together with minimal false positives. The evaluation metrics demonstrate strong signs of training overfitting since both the validation accuracy and the recall suddenly change during different training stages. The validation accuracy decreases at epochs 6, 7, and 15, which creates three points of poor generalization during training. Both critical parameter adjustments and additional regularization methods prove necessary to achieve the optimization progress. For improved model performance, practitioners should adopt either an optimized learning rate schedule or metaheuristic approaches that optimize adjustable hyperparameters. The basic model achieves optimal performance across every phase of measurement, with 98.09% accuracy and 0.999 AUC. Applying GA optimization methods would allow users to adjust hyperparameters better so that they can extend their generalization abilities and reduce performance fluctuations.

4.2 Model Evaluation With the Genetic Algorithm

The development process of the proposed DL model with the GA incorporated utilized accuracy, loss, precision, recall, and F1 score, combined with AUC metrics for assessment. The GA conducts automated hyperparameter optimization, through which the model can adjust its learning rates in addition to managing the dropout values and dense layer unit values. The enhanced model provided two essential benefits: stabilized training operations and improved generalized function output. The training process stretched across 50 epochs, as shown in Figure 9, which shows (a) accuracy, (b) loss, (c) precision, (d) recall, (e) F1 score, and (f) ROC curve performance metrics. The GA-optimized model outperformed its nonoptimized counterpart on the basis of Figure 9a. During the initial epoch, the model had an accuracy of 38.93%, but it reached 96.79% accuracy in the final epoch. The model achieved 90.21% validation accuracy after demonstrating a continuously increasing trend during its last epoch. The GA-optimized model parameters lead to better model classification results because of their effectiveness in optimization. The training loss function substantially decreased, according to Figure 9b. The model initiated with a training loss of 0.067 before decreasing to 0.002, thus demonstrating adequate learning and error reduction. During hyperparameter optimization employing the GA, the validation loss achieved a consistent downward trend starting at 0.035 and ending at 0.009, whereby this outcome established enhanced generalizability in the model. Figure 9c shows how the model performed by correctly identifying positive results through its precision curve assessment.

The model achieved a training precision rate of 98.76% when moving from 47.42% at Epoch 1 to Epoch 50. The GA proved effective in finding optimal hyperparameters because its consistent improvements demonstrated better decision-making capabilities for the model. The validation precision displayed a regular upward trend, which reached 93.67% in the final epoch. Figure 9d shows the recall level based on model effectiveness in identifying actual positives. The initial training recall was 27.31%, but it rose to 94.54% after the last training epoch. During Epoch 50, the validation recall metric reached 86.25%, which proved that the model acquired a better ability to detect every target occurrence. The F1-score curve in Figure 9e demonstrates the model's robustness because it calculates precision and recalls values as the harmonic mean. The F1 score values increased progressively throughout training due to the enhanced precision–recall performance. The iterative optimization through the GA drove an increase in the validation F1 score since it refined the balance between the detection accuracy and the number of false results across successive training cycles. This figure shows the discriminative ability assessment of our model through the ROC curve. The AUC value increased from 0.722 at the beginning of Epoch 1 to 0.999 by Epoch 50, which demonstrates the enhanced classification ability of the GA-optimized model. Higher AUC scores identify the model, as it achieves strong class discrimination along with minimal false-positive occurrences. The GA enables important optimization of hyperparameters to achieve better performance, generalization, and stability in the model system. Several experiments demonstrated that the model achieved 96.79% accuracy, which was supported by the AUC value of 0.999, thus proving its capacity for complex classification operations. Metaheuristic optimization technology is effective for DL models, enabling their better application in real-world scenarios.

4.3 Confusion Matrix

The confusion matrix analyzes model classification performance by showing how many instances the model correctly classified as well as its incorrect identifications for each category (Figure 10) (Rai and Pahuja 2023). The confusion matrix of the nonoptimized model reveals the prediction distribution across the six disease classes. Overall, the model demonstrates reasonably strong performance, correctly classifying the majority of samples. For the aphid class, the model achieved 38 correct predictions out of 40 samples, misclassifying one image each as bacterial blight and target spot. All 40 samples of army worms were accurately identified, indicating strong feature differentiation for this class. However, bacterial blight exhibited some classification difficulty, with 36 correct predictions and three misclassifications, one as aphids and two as target spots. The healthy leaf category achieved 38 accurate predictions but suffered two misclassifications into bacterial blight. Similarly, the model recognized 36 powdery mildew instances correctly but made one mistake each in identifying them as aphids, healthy leaves, and target spots. Finally, the target spot had 39 correct classifications and one error, misclassified as bacterial blight. Although the model performed well overall, it showed challenges in distinguishing between visually similar diseases like bacterial blight and target spot, which often exhibit comparable lesion characteristics. The classification errors suggest that the model, without GA-based optimization, may lack sufficient tuning in terms of dropout rate, learning rate, and hidden layer configuration, limiting its ability to extract and generalize subtle discriminative features. As a result, certain minority or symptomatically overlapping classes remain prone to confusion. Incorporating an intelligent optimization strategy like the GA is essential to enhance model sensitivity, feature discrimination, and overall classification accuracy.

The classification output of the optimized model, trained using GA-based hyperparameter tuning, is shown in Figure 11 through an updated confusion matrix. This matrix reveals improved classification stability and accuracy across all disease categories as a result of refined tuning of learning rate, dropout rate, and dense layer neuron count. The model correctly identified 38 aphid samples out of 40, misclassifying one as a healthy leaf and another as a target spot. All 40 army worm samples were once again correctly classified, maintaining the perfect performance achieved by the baseline model. The bacterial blight category showed significant improvement, achieving perfect classification with all 40 samples correctly labeled, whereas the non-GA model had previously struggled with this class. Under the healthy leaf category, the model was able to accurately predict 38 samples of 40, and the other two samples were misclassified, which were in cases of bacterial blight. Powdery mildew worked better as well, since it was 38 correctly identified and only two errors, one mistake concerning bacterial blight and one concerning target spot. Lastly, the target spot class had a positive prediction of 39, and one sample that was turned into bacterial blight. The percent misclassifications were much lower and of a more distributed pattern in the GA-enhanced model than in the baseline. Such advances are even more visible in the categories of bacterial blight and powdery mildew, which were problematic ones before. The improved classification accuracy has been due to the experimental capability of the GA to complement the hyperparameter space, finding the configurations that are good to lead to enhanced convergence and pattern learning. GA optimization has the effect of encouraging the frequent regularization and good feature transformation in the dense layers, which minimizes the confusion between visually semantic classes, enhancing the model generally in terms of its generalization ability. The confusion matrix optimized delivers the visual confirmation of the success of the GA in lifting the diagnostic quality and deployment readiness of the model, which will help in real-life cotton disease identifications.

The above misclassifications between bacterial blight and target spot can also be attributed to visual similarity in the patterns of their lesions, especially when they reach almost the same stages of development. Both conditions have dark lesions of necrosis or darkened areas of chlorosis, appearing morphologically similar in RGB images with dimmable light, or when leaves are old. This phenotypic similarity is a burden even to human experts, and it may lead to predictions of two adjacent models. Future enhancements can contain domain-specific data augmentation, like lesion-specific average patch cropping, normalization of colors, or image variations to focus on lesion forms and texture variations. Also, there may be value in feeding multispectral imaging data or a sequence of disease progression into the model and training to learn the subtler temporal or spectral differences between these classes that may exist. Such improvements would elevate class separability and model robustness in the field.

4.4 Hardware Setup and Efficiency

The model training and testing were performed with a 2022 MacBook Air with Apple M2 chip (8-core processor, 10-core graphics), unified memory 8 GB, and operating system macOS Sequoia 15.3.2. The training process took an average of approximately 58 min, despite the lightweight nature of this device, leading to an effective training process of 50 epochs. The architecture of the model employed the EfficientNetB3 and InceptionResNetV2 backbones, which were already trained, thereby minimizing the computational burden and spurring convergence. In the evaluation process, the inference times on average per image were noted at 32 milliseconds, indicating that the model can provide near-real-time predictions. These findings show that the suggested method is computationally cost-effective and aligns with application settings, where reaction time is also a critical factor in the course of implementation, for example, during field-based diagnosis of practically important disease types, in a mobile implementation of applications.

The relatively short training time of 58 min for the proposed hybrid CNN-GA model can be attributed to a combination of optimization strategies. First, we employed transfer learning by using ImageNet-pretrained EfficientNetB3 and InceptionResNetV2 models with their convolutional layers frozen during training, thereby reducing the number of trainable parameters and computational overhead. Second, the feature extraction phase was performed once and reused throughout the GA optimization process, eliminating the need to retrain the full CNNs repeatedly. Third, the final classification model, optimized via GA, was a lightweight, fully connected network trained on concatenated feature vectors, which significantly reduced training complexity. Lastly, the use of efficient data generators and TensorFlow's support for hardware-level acceleration and mixed-precision arithmetic contributed to faster training convergence. These combined strategies enabled us to achieve high model performance with reduced computation time, even on a resource-constrained M2 MacBook Air.

4.5 Performance Parameters

The researchers provided a performance analysis showing the comparison results between models built with and without GA integration through Table 2 using several classification metrics, including accuracy and precision, recall, F1 score, and AUC (Borugadda et al. 2021). The classification performance metrics reveal the effectiveness levels for model discrimination between diseased and healthy cotton leaves. Each class in the non-GA model demonstrated contrasting precision, recall, and F1 score values, leading to 90.21% model accuracy. The model performed well in class detection according to its high AUC values but displayed reduced recall values for various disease categories, which indicated potential diagnostic errors. This performance demonstrates a discrepancy in identification capability since some classes receive better recognition outcomes than others do. The model optimized through the GA reached 92.45% accuracy while demonstrating higher precision measures alongside improved recall and F1 scores throughout all the experiments. Through adaptive hyperparameter tuning controlled by the GA, the model acquired a stronger ability to discover complex relationships between features, which helped it distinguish visually similar leaf conditions. An analysis of the recall values proved significant because the GA-assisted model achieved improved results in terms of recognizing actual diseased samples, which reduced the number of false negative callbacks. The GA optimization led to better disease category discrimination through a slight increase in the AUC values. The results indicate that the GA succeeded in reducing overfitting, so the model maintained strong performance in new test data situations. The data in Table 2 establish that the GA-assisted model provides superior classification reliability and robustness compared with the non-GA model structure.

4.6 Results of LIME and SHAP Visualization

The interpretability of models is enhanced through LIME, which shows significant features that impact predictions. The system creates disruptions in the input data to examine prediction alterations for generating local explanations. The application of LIME for cotton leaf disease classification is presented in Figure 12. The model has the strongest confidence in Class_5 according to the prediction probabilities shown in Figure 12a. The important feature values behind the prediction include Feature_978 and Feature_1412, as shown in Figure 12b. The local explanation in Figure 12c provides a visual depiction of feature contributions by using bar plots to display both positive and negative values. Figure 12d offers visual explanations that demonstrate that green areas indicate positive effects on classification, but orange-red areas lead to negative prediction results. Through the LIME approach, organizations gain confidence due to model transparency, which enables users to identify errors and make necessary adjustments. The requirement for explainable AI systems is fundamental for implementing AI-based cotton leaf disease detection systems in operational agricultural environments.

To address the interpretability concerns and strengthen the explainability of the model, additional LIME visualizations were generated for multiple representative images from each disease class. These visualizations, now included in an extended version of Figure 12d, highlight the most influential pixel regions contributing to the classification decision. The LIME results show that the model often focuses on disease-specific lesion patterns, discoloration zones, and vein distortions, features that align with those used by human agronomists during manual inspection. For instance, in the case of bacterial blight, the model concentrated on the water-soaked lesion edges, while for powdery mildew, it localized to white fungal patches. These findings indicate that the model uses biologically relevant cues for decision-making. Additionally, in a few cases, the model attended to noninformative background textures, suggesting minor overfitting, which could be mitigated through targeted data augmentation or transfer learning with attention guidance. These insights not only validate model decisions but also provide a roadmap for future training refinements and expert-aligned learning.

The cutting-edge explainability method known as SHAPs involves users whose model features make a difference in the prediction outcome. SHAP uses Shapley value computation to determine how each feature element affects the model prediction outcome. Information about the key parameters influencing different cotton leaf diseases can be found in the SHAP summary plot presented in Figure 13. Each input feature receives a measurement of importance on the basis of the assessment of mean (|SHAP value|) scores, which reflect its influence level on model predictions. The visualization reveals three main features, namely, Feature 570, Feature 956, and Feature 674, because they strongly affect the classification results, which target Class 3 red and Class 5 purple. The visual representation based on color provides quick comprehension of which characteristics play a role in which disease groups. The model's interpretability gives experts the power to understand the DL system's ability to detect diseased cotton leaves, thus enhancing their confidence in automated disease recognition solutions.

Figure 14 presents SHAP summary plots for each of the six disease classes: (a) Class 0 through (f) Class 5. These plots display the top features influencing the model's prediction for each class. Each point represents an individual prediction instance, where the color indicates the feature value (red for high, blue for low), and the horizontal position indicates the SHAP value. This visualization aids in understanding class-specific feature impact and model behavior.