MCCLDP: Multi Class Cotton Leaf Diseases Prediction and Classification using Deep Learning Model
Kokkula Shiva Prasad, A. Balaram, Nagunuri Rajender, P. Sudheer, Sneha Yerram, Sumayyam Begum
Abstract
Cotton plant disease detection is critical for sustainable agriculture and reducing crop losses. This paper proposes a novel Multi-Stream Attention-Guided Hybrid CNN (MAH-CNN) for accurate classification of cotton leaf diseases. The model leverages pre-trained ResNet152v2 and DenseNet-121 backbones for hierarchical feature extraction, complemented by a shallow CNN for localized texture analysis. A spatial attention mechanism enhances focus on disease-relevant regions, mitigating background noise. Features from the global and local streams are fused and passed through a lightweight classification head. The model achieves superior performance in terms of accuracy 97.32%, F1 score 98%, and specificity 100% on benchmark datasets which are available in open access, outperforming existing state-of-the-art methods. The integration of Grad-CAM provides interpretability, fostering trust in automated disease detection systems.