Self-supervised deep learning spots cotton leaf diseases with explainable heatmaps

Cotton, often called “white gold,” loses quality and yield to a range of leaf diseases, including Alternaria leaf spot, bacterial blight, curl virus and other stress-related conditions. Diagnosing these problems traditionally depends on expert knowledge and manual inspection, a process the authors describe as labor-intensive, subjective and error-prone. A new study in Neural Computing and Applications presents a deep learning framework that aims to automate this diagnosis while requiring fewer labeled images and explaining its own decisions.

The work, published on 20 October 2026 as Volume 38, article number 827, was carried out by Anjali Singh, Padma Nilesh Mishra and Garima Singh. Their approach combines three ingredients: self-supervised learning (SSL) for pretraining, a supervised convolutional neural network (CNN) for classification, and an explainable AI technique called Gradient-weighted Class Activation Mapping, or Grad-CAM.

The central problem the researchers set out to address is the heavy dependence of most current plant-disease models on large labeled datasets.

To reduce that dependence, the framework first applies self-supervised contrastive pretraining to a set of unlabelled cotton leaf images. In this stage the network learns general and distinctive feature representations without any disease labels, essentially teaching itself what visually distinguishes one leaf from another. These pretrained representations are then fine-tuned with labeled data to classify multiple types of cotton leaf diseases.

Transparency is handled by Grad-CAM, which generates visual heatmaps overlaid on the leaf images. These heatmaps highlight the areas affected by disease, allowing domain experts to check whether the model is attending to biologically meaningful regions rather than background artifacts. According to the authors, the visualizations confirm that the model targets biologically relevant areas, which they argue boosts understanding and user trust.

The framework was evaluated on the SAR-CLD-2024 cotton leaf dataset, which contains 2,137 labeled images spanning seven classes, plus a separate unlabeled subset of 565 images reserved for self-supervised pretraining. The dataset deliberately reflects the messiness of real-world agriculture, showing class imbalance, natural variability and real-field imaging conditions rather than clean laboratory photographs.

The reported results are strong. The framework reaches a validation accuracy of up to 94.38 percent and achieves strong precision, recall and F1-scores across all disease classes. The authors state that this performance exceeds that of several leading supervised and transfer learning models they compared against.

Beyond raw accuracy, the study’s broader claim is practical: by pairing self-supervised pretraining with explainable outputs, the framework lowers the need for large labeled datasets, increases robustness and generalization, and offers what the authors call a practical solution for accurate and understandable cotton leaf disease detection in real-world farming situations.

As with any single-dataset evaluation, however, performance on other cotton varieties, regions and imaging setups remains to be demonstrated, and the reported accuracy reflects validation on this particular benchmark rather than an independent field deployment.

Subject of Research: Self-supervised deep learning with Grad-CAM explainability for detecting cotton leaf diseases

Article Title: Deep learning-based detection of cotton leaf diseases using self-supervised feature learning and Grad-CAM visualization

Article References: Singh, A., Mishra, P. N., & Singh, G. (2026). Deep learning-based detection of cotton leaf diseases using self-supervised feature learning and Grad-CAM visualization. Neural Computing and Applications, 38(20), Article 827. https://doi.org/10.1007/s00521-026-12586-y

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12586-y

Keywords: deep learning, cotton leaf diseases, self-supervised learning, Grad-CAM, explainable AI, convolutional neural networks, plant disease detection, SAR-CLD-2024, agriculture, contrastive pretraining, Deep, learning-based