Deep Learning Model Reaches 98.39% Accuracy in Cotton Leaf Disease Classification

Researchers report that a deep learning framework combining the ConvNeXt convolutional architecture with a modified swarm intelligence algorithm called the Random Walk Grey Wolf Optimizer (RWGWO) classified cotton leaf diseases with 98.39% accuracy on a curated public image dataset, outperforming convolutional neural network, self-organizing map artificial neural network, and support vector machine baselines tested under identical experimental conditions. The study, published open access in Discover Artificial Intelligence on 23 October 2026 by Yuvraj Wagh, Ashwin R. Dobariya, and Ashish Kumar Nayak, addressed six categories: aphid infestation, armyworm damage, bacterial blight, powdery mildew, target spot, and healthy leaves. The authors position the work as a step toward automated crop monitoring, while explicitly acknowledging that their results come from a laboratory-style benchmark rather than independently collected field data.

The motivation is grounded in the economics of cotton farming. India has the largest cotton-cultivated area in the world, and the authors identify leaf diseases as one of the most significant causes of yield loss in the crop. Investigations conducted between 2017 and 2019 attributed substantial cotton production losses to pest infestation and adverse weather patterns. Late detection compounds the problem, since treatment costs rise and diseases spread, and the paper notes that relatively few agricultural experts are available for farmers to consult at an early stage. The five disease categories studied here all disrupt the plant’s physiological systems and vascular function, reducing both yield and fiber quality, which is why the authors argue that automated monitoring of large cultivated fields could reduce economic losses and support more sustainable cotton production.

ConvNeXt, the backbone of the proposed system, is not a transformer despite its transformer-era pedigree. The authors emphasize that it is a convolutional neural network redesigned with modern features borrowed from the vision transformer literature, including large kernel sizes, inverted bottlenecks, layer normalization, and GELU activations, arranged in a deeper ResNet-like structure. This design places it, structurally, between classical CNNs and architectures such as Swin Transformers and EfficientNetV2, while retaining the inductive bias and computational efficiency of convolutions. The researchers applied transfer learning, initializing from models pretrained on the ImageNet-22k dataset and then fine-tuning on the cotton leaf data, a strategy intended to help the network detect fine-grained disease patterns despite the modest size of the target dataset.

The second half of the hybrid, RWGWO, is an extended version of the Grey Wolf Optimizer, a metaheuristic that mimics the leadership hierarchy of a wolf pack by tracking the three best candidate solutions, conventionally labeled alpha, beta, and delta, and averaging their guidance to update the rest of the population. Coefficient vectors control the balance between exploration and exploitation as a control parameter shrinks over the course of the search. The random walk extension addresses a well-known weakness of swarm methods: stagnation. When fitness fails to improve for a set number of consecutive iterations, the algorithm perturbs each candidate position with a Cauchy-distributed random step, with the step size constrained between 0.01 and 0.1. Because the Cauchy distribution is heavy-tailed, it permits occasional long jumps that can carry the search out of local optima. Importantly, the authors clarify that RWGWO operates as an outer-loop hyperparameter optimizer around ConvNeXt training; it does not alter the network’s internal convolutional blocks or replace backpropagation, so the reported gains stem from better hyperparameter selection and search-space exploration rather than architectural modification.

Before any of this optimization, the image pipeline underwent extensive preprocessing. Each picture was segmented, and pixel outliers were filtered using Gaussian, median, and wavelet filters implemented through the PyWavelets library, followed by histogram equalization to boost contrast. All images were resized to 224 by 224 pixels and normalized using the dataset’s mean and variance. Data augmentation, including horizontal and vertical flips and rotations of 60 and 120 degrees, was applied to maintain balanced class representation and reduce overfitting. The dataset, drawn from a public Kaggle cotton plant disease collection, contained 3,601 images split almost evenly across the six classes, roughly 600 per category. Training used 70% of the data, with 15% each for validation and testing, allocated through stratified sampling so that every class was proportionally represented, and a randomization seed of 42 was fixed for reproducibility. Layer freezing, early stopping, and L2 regularization were incorporated to curb overfitting, and the ablation study found that removing the Cauchy-distributed perturbation cost 1.3 percentage points of F1-score.

The headline result is a test-set classification accuracy of 98.39% for the RWGWO-optimized ConvNeXt, compared with 96.74% for a conventional CNN, 92.00% for a self-organizing map artificial neural network, and 90.00% for a support vector machine. Statistical analysis reported p-values below 0.01 for the accuracy comparisons against these baselines, alongside 95% confidence intervals, and ConvNeXt showed lower variance across multiple independent training runs, which the authors interpret as greater stability and reduced sensitivity to particular test splits. Receiver operating characteristic analysis yielded area-under-the-curve values between 0.93 and 0.98 for most classes. Per-class F1-scores ranged from 99% for healthy leaves, the easiest category, down to 96% for bacterial blight, whose symptoms overlap with other stress factors. The authors also report precision of 97% and recall of 99% for the model overall, figures they note indicate few false positives and false negatives, outcomes that matter in agriculture because a false positive can trigger unnecessary chemical spraying while a false negative can allow a disease outbreak to progress unchecked.

The error analysis is arguably the most practically informative part of the paper. The normalized confusion matrix shows that the model most often confused aphid-infested leaves with bacterial blight, two conditions whose early symptoms can be subtle enough to deceive even trained human observers. Armyworm damage posed a different challenge: its small lesions resemble other leaf diseases in early stages, and the corresponding AUC values were comparatively lower than for fungal or bacterial diseases, which tend to produce more prominent lesions. The authors stress that low recall on armyworm is particularly dangerous in deployment, because underestimating an insect population can lead to missed pest control and substantial crop loss, while the opposite error, flagging healthy leaves as diseased, wastes pesticides and money. They suggest that class weighting or threshold adjustment along the precision-recall curve could improve recall on minority classes such as armyworm without sacrificing overall performance.

To test whether the random walk component itself deserved credit, the team compared RWGWO against two established gradient-based optimizers, AdamW and AdaBelief, on the same dataset. RWGWO consistently edged out both on accuracy, precision, and recall, with the gap most pronounced for correctly identifying diseased leaves, and it achieved the highest F1-score of the three, while AdaBelief and AdamW recorded equal F1-scores but struggled more with minority classes. The trade-off is computational: RWGWO consumed more resources and converged more slowly than the gradient-based alternatives, reflecting its wider exploration of the hyperparameter space, though the authors judge the accuracy and generalization gains worth that cost. They also acknowledge a caveat about attribution. Because the comparison covered only AdamW and AdaBelief, it does not constitute a full metaheuristic ablation; controlled experiments against standard Grey Wolf Optimizer and population-based methods such as particle swarm optimization or the Whale Optimization Algorithm, under equal search budgets, have not been performed, so no numerical improvement is claimed for those missing configurations.

The authors are unusually candid about the limits of their benchmark. The Kaggle collection lacks the variability of real fields, including inconsistent lighting, cluttered backgrounds, occlusion, weather effects, and device variation, and original camera models, geographic locations, acquisition dates, and a formal expert-annotation protocol were not available in the experimental records. No separate field cohort was collected, and the paper states plainly that the 98.39% figure cannot be regarded as the model’s actual generalization performance in the field. Formal k-fold cross-validation was also omitted, primarily because repeatedly training the ConvNeXt-plus-RWGWO pipeline is computationally expensive, though multiple independent runs were used to estimate variance. On the deployment side, parameter counts across the ConvNeXt family range from 8.6 million for ConvNeXt-Tiny to 366.4 million for ConvNeXt-XLarge, and the larger variants’ memory and energy demands make them poorly suited to edge devices. The authors identify ConvNeXt-Tiny and ConvNeXt-Small as the practical candidates for mobile diagnostic tools, while noting that per-image inference latency and memory footprint on representative edge hardware were not benchmarked in this study.

Future work outlined in the paper spans several directions: validating the model on independently collected field datasets across geographic and climatic conditions and on more cotton strains and disease types; adding explainable AI methods such as SHAP or LIME-based visual attribution to reveal which leaf regions drive each prediction; packaging the model into web or mobile applications; integrating live feeds from IoT-connected cameras or drones for continuous monitoring; and extending to multi-modal learning that incorporates environmental variables such as humidity, temperature, and soil condition. If those steps bear out the laboratory results, the authors argue, early automated detection could reduce pesticide waste tied to late diagnosis and strengthen the textile fiber supply chain. For now, the study stands as a carefully documented demonstration that a transformer-influenced CNN, steered by a stagnation-resistant metaheuristic, can set a strong benchmark on cotton leaf disease classification, with its real-world value still awaiting field trial.

Subject of Research: Deep learning classification of cotton leaf diseases using ConvNeXt with Random Walk Grey Wolf Optimizer hyperparameter tuning

Article Title: Enhanced cotton leaf disease classification Through ConvNeXt and Random Walk Grey Wolf Optimization

Article References: Wagh, Y., Dobariya, A. R., & Nayak, A. K. (2026). Enhanced cotton leaf disease classification Through ConvNeXt and Random Walk Grey Wolf Optimization. Discover Artificial Intelligence, 6(1), Article 1599. https://doi.org/10.1007/s44163-026-02566-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02566-5

Keywords: cotton leaf disease, ConvNeXt, Grey Wolf Optimizer, deep learning, plant disease detection, precision agriculture, hyperparameter optimization, transfer learning, computer vision, crop protection, machine learning, agricultural AI