Abstract

Bacterial blight, curl virus, and leaf reddening are cotton leaf diseases that present a serious threat to the crop yield and fiber quality, necessitating the need to diagnose them correctly and on time. The manual inspection is tedious, subjective, and cannot be applied to large-scale farming, which is why automated systems of deep learning-based detection were developed. In general, current methods are usually highly accurate but fail to be robust with class imbalance, statistical reliability, interpretability, and computational efficiency, which restrict real-world applications. To overcome these issues, we suggest an effective and explainable deep learning network based on a Squeeze-and-Excitation (SE) improved DenseNet201 to classify cotton leaf disease in multiple classes. The model integrates SE blocks for adaptive channel-wise feature recalibration, robust preprocessing and data augmentation for class imbalance, and employs Grad-CAM and LIME for global and instance-level interpretability. Performance is assessed with macro-averaged measures, measures of reliability and five-fold cross-validation. The experimental findings show that the proposed SE-DenseNet201 attains a test accuracy of 98.64%, a macro F1-score of 0.9968, and a Cohen Kappa of 0.9836, and consistent five-fold cross-validation scores. This model has a practical inference time of 0.96 seconds per image and the average computational demands (GPU: 316.92 MB, RAM: 1.97 GB) whereas Grad-Cam and LIME visualization attests diseaserelevant features attention. Such results indicate that the framework proposed is capable of providing accurate, reliable and interpretable predictions, which would provide a viable solution to automated cotton leaf disease detection and enable precision agriculture applications.