AGAComba: Adaptive Graph Attention and Comba for Rice Leaf Disease Diagnosis

Rice leaf diseases are a global issue. There are numerous types of leaf diseases, and their control requires targeted measures. These diseases directly affect the growth and yield of rice. Using deep learning methods for diagnosing rice leaf diseases is an extremely effective approach. However, the existing methods usually capture the characteristics of leaf diseases in a static manner, and they encounter difficulties in overcoming the imbalance and multi-scale issues when dealing with complex background rice images. We propose a new model that combines Adaptive Graph Attention (AGA) with Comba, and call it AGAComba. The AGA module captures key information through adaptive feature calibration. The AGAComba model employs a closed-loop control to enhance the processing of long-term sequence information and their dependencies, improving the model’s understanding of previous information and optimizing the diagnostic results. Experiments on a comprehensive dataset of rice leaf diseases demonstrated that AGAComba has outstanding performance, with a test accuracy rate of 95.35%.