COLD-12: A multi-level feature extraction hybrid CNN Model for accurate cotton disease diagnosis

Cotton leaf diseases are common agricultural issues and multifaceted threats to millions of farmers worldwide. This study aims to provide an up-to-date approach to diagnosing diseases of cotton leaves using advanced DL (Deep Learning) that is accessible and actionable by farmers, bridging the gap between technology and agricultural practices. The key to this study was the construction of a diverse dataset comprising 5722 high-resolution images collected from three independent sources and further enriched with expert guidance in annotation. Advanced preprocessing was done, like denoising, sharpening, and color balancing, to increase the quality of images, while augmentation increased generalization and reduced overfitting. In it, COLD-12 is proposed, which has achieved a promising result by incorporating Atrous Spatial Pyramid Pooling with Squeeze-and-Excitation blocks for improved channel attention in its multi-level feature extraction. The training accuracy of the model was 99.94 %, and the validation accuracy was 99.24 %, which indicates that the model was robust with very low validation loss. The inner workings of the COLD-12 model were interpreted by using explainable AI techniques such as Grad-CAM, Grad-CAM++, and Layer-CAM. It showed advantages compared to state-of-the-art models such as VGG16, VGG19, Xception, DenseNet121, and InceptionResNetV2.The impact of batch size, K-fold cross-validation, and preprocessing techniques was also explored, resulting in improved accuracy and interpretability. An interactive web tool was developed to bridge research and real-world applications, offering farmers a convenient assistant for diagnosing cotton leaf diseases, thus supporting sustainable farming initiatives.