A Hybrid Deep Wavelet Feature Fusion Framework for Rice Leaf Disease Classification

Abstract

Accurate identification of rice leaf diseases is crucial for minimising yield loss and enabling effective crop management. Although convolutional neural network (CNN)-based classification approaches have demonstrated considerable performance in recent years, their successive downsampling operations lead to the loss of fine-grained spatial details that are essential for detecting subtle disease patterns. To overcome this limitation, this study proposes a hybrid feature extraction framework that integrates deep representations with handcrafted wavelet-based descriptors within an end-to-end learning paradigm. While the CNN model captures complex, high-level semantic patterns, the wavelet-based component extracts fine-grained texture characteristics and domain-specific structural variations, including edge details and localised intensity changes that are often attenuated during convolutional and pooling operations. The fused feature representation is subsequently fed into a classification head for final disease prediction. In addition, a dataset comprising four prominent rice diseases such as brown spot, bacterial leaf streak, rice blast, and sheath blight specific to the Northern Province of Sri Lanka was constructed in this study. Experimental evaluation on five benchmark datasets demonstrated the effectiveness of the proposed approach, achieving an accuracy of 95.07% on the self-constructed dataset.