Spectral spatial transformer with spatio-temporal decoupled attention for rice fertilizer efficiency classification

Rice is a globally important staple crop, and efficient fertilization management is crucial for agricultural sustainability and food security. Traditional experience-based methods fail to meet the dynamic monitoring needs of precision agriculture, while existing deep learning models often suffer from dimensional mismatches, weak spatial feature extraction, and overlooked temporal information, leading to reduced classification accuracy. To address these limitations, this paper proposes an S2TDA architecture based on a spatial-temporal decoupled attention mechanism. During spatial encoding, convolutional and channel attention modules are incorporated to adaptively weight six-channel multispectral inputs and suppress redundant bands. A residual-normalized spatio self-attention module is further introduced to enhance local texture extraction. Moreover, a spatio-temporal decoupled attention mechanism is designed to separately learn spatial and temporal dependencies by performing temporal self-attention on features extracted from each phase, enabling effective modeling of crop growth dynamics. Experimental results demonstrate that the improved architecture significantly enhances classification accuracy for multi-temporal multispectral images. Compared to the original framework, the improved approach achieves approximately 35% higher overall and average accuracy, with the Kappa coefficient increasing from 0.41 to 0.79.