Dual-Interaction Spatiotemporal Fusion Network for Agricultural Remote Sensing Imagery

Spatiotemporal fusion (STF) enables the integration of remote sensing (RS) images with high spatial and temporal resolutions, playing a vital role in agricultural applications such as crop monitoring and disease/pest early warning. Recently, Transformer has been applied in the field of STF and has aroused widespread attention due to its powerful representation ability for capturing global long-term dependence features. However, the Transformer applied to STF utilizes self-attention with spatial windows, coupled with channel-wise weight sharing, which significantly constrains cross-spectral adaptive modeling. In addition, conventional feature-extraction frameworks exhibit substantial attenuation of high-frequency components, resulting in spatially blurred fusion outputs. In this study, we propose a Dual-Interaction Network (DINet) for STF of agricultural RS. The DINet operates on two cross-paired inputs (a fine reference image and a coarse target image), significantly relaxing the strict requirements of conventional STF methods. In the proposed framework, a spatial-channel Transformer (SCT) is developed to adjust receptive fields in light of the distinct resolutions of coarse/fine inputs, enhancing spatial-channel feature interaction during the feature-extraction stage. A wavelet decomposition-enhanced strategy is leveraged to integrate high-frequency details and low-frequency global information through a spatial-temporal Transformer (STT). In the reconstruction stage, a multi-level fusion module (MLFM) is designed to hierarchically fuse the features from multiple branches, boosting the spatial-temporal consistency. Experimental results demonstrate that the proposed method outperforms state-of-the-art approaches in preserving fine-grained agricultural features while maintaining temporal consistency, achieving superior performance for crop monitoring tasks.