DenseNet169-Enabled High-Accuracy Automated Detection System for Cotton Leaf Diseases Using Deep Transfer Learning

Abstract

Rapid and accurate identification of cotton foliar diseases which seriously threatens cotton output worldwide is of great importance. A strong deep learning model with custom architecture DenseNet169 is proposed in this research for automatic classification of seven diseases of cotton leaf: Bacterial Blight, Curl Virus, healthy leaf, herbicide growth damage, leaf hopper jassids, leaf redding and leaf variegation. We proposed a two-step transfer learning method with enhanced data augmentation based on the SAR-CLD–2024 dataset, which contains 9,137 images. The DenseNet169 architecture proposed here yielded a remarkable performance with a validation accuracy of 96.83% while precision, recall, and F1-score were 96.89%, 96.93%, and 96.90%, respectively, with a significant enhancement than prior related approaches. It gets flawless classification for Herbicide Growth Damage and close to perfect for each disease types with macro-average AUC of 99.81%. The second is the parameters of the deep architecture that we adapted for agricultural pathology where we were broadly successful at systematic feature extraction and then fine-tuning 161 layers. The new high-water mark physiological plant disease diagnosis that we establish here is an important step toward ultimately real-world applicability as adaptive components of precision agriculture to monitor crop health and protect yield.