Edge-Guided Multiscale Spatio–Temporal–Spectral Dynamic Coordination for Agricultural Parcel Classification

Satellite image time series (SITS) has become a preferred data source for precise farmland mapping. However, current deep learning models are mainly relied upon to automatically mine temporal, spatial, or spectral dimensional features from static inputs, ignoring the dynamic interactions between these dimensions as well as information about the structured boundaries of agricultural plots. In order to effectively utilize the multidimensional information embedded in SITS data and improve the classification accuracy of time-series crops, we constructed a one-to-one dynamic transform based on the characteristics of spatio–temporal–spectral information. Meanwhile, we used an edge detection algorithm to extract the coordinates of edge pixels, dynamically generated binary masks as the edge a priori, and combined the residual network and self-attention to propose a U-shaped encoder–decoder based on the dynamic mapping of features classification network called TSSDynamicNet. Our experiments, conducted on SITS data from Sutter and Kings counties, California, USA, demonstrate that the proposed method consistently outperforms representative deep learning approaches, achieving state-of-the-art performance.