Abstract
Accurate cropland change detection in very-high-resolution (VHR) imagery requires synergizing fine-grained spatial details with long-term temporal dynamics. We propose the two-stage synergistic temporal network (TS-SynT), a novel framework that decouples spatial representation learning from temporal trend analysis. The first stage, the synergistic spatial enhancement system, leverages dual VGG16 and FastSAM backbones with three modules: the feature context builder module (FCBM) for multiscale context, the gated synergy module for spatio-temporal distortion correction, and the enhanced attention optimization module for boundary refinement. The second stage, the spatially conditioned temporal mixer, utilizes high-quality spatial priors to interpret long-term NDVI sequences, distinguishing fundamental land-use changes from transient disturbances (e.g., phenological shifts). Comprehensive sensitivity and ablation evaluations demonstrate TS-SynT’s state-of-the-art performance; notably, on the highly challenging and multitemporal SW-CLCD-TS dataset, TS-SynT achieves a peak F1-score of 72.66% and an intersection over union of 57.05%, substantially outperforming both uncoupled postclassification sequential baselines and alternative sequence modeling backbones. This work advances intelligent agricultural monitoring, supporting global food security and sustainable land management.