Transformer-Embedded Attentive CNN for Spectral Image Analysis of Rice Blast Syndromes

Abstract

Leaf blast disease is a significant constraint in world-wide rice production systems, necessitating effective monitoring for optimized crop-yield management. Satellite-derived land-surface temperature data can be an essential input for detecting such a disease, as it plays a critical role in the pathogen’s development and spread. When combined with other environmental factors, such as humidity and leaf wetness, it serves as a key indicator of potential outbreaks. Vegetation and moisture indexes captured by the European Space Agency satellite Sentinel 2, have been used to analyze rice blast disease on a large scale. However, due to substantial geo-spatial and temporal variability, predicting disease occurrence remains a challenge. To address this gap, we propose a fusion of convolutional neural networks (CNN) and transformer-based models to reveal both local and global syndromes in images associated with the risk of rice blast disease. A novel multichannel attention mechanism within the CNN helps extract essential spectral information, where each RGB channel’s spatial intensity is leveraged to focus on critical details through multihead attention. The transformer network with dynamic tokenization and self-attention captures global information, enabling lightweight transformers to highlight discriminative global features. Dynamic tokenization selects tokens or patches based on attention factors, facilitating the extraction of important sequential information. The aggregated network output enhances the classification accuracy of leaf blast risk prediction while reducing computational complexity. The proposed approach outperforms existing models in spectral image analysis for predicting the spread of leaf blast disease.