Answers are grounded in this paper and cite their sources.
Farmland change detection is crucial for ensuring food security and sustainable agricultural management. While deep learning has shown promise, its application is severely hampered by the scarcity of high-quality labeled data in practical agricultural scenarios. To address the challenges of poor feature generalization and inadequate spatio-temporal modeling in unsupervised settings, this paper proposes STCU (Spatio-Temporal Change Understanding), a novel unsupervised framework that synergizes a domain-adapted Segment Anything Model (SAM) with advanced temporal modeling. The core contributions are threefold: First, we introduce Low-rank Adaptation (LoRA) to fine-tune the pre-trained SAM, effectively mitigating the domain discrepancy between natural and remote sensing images. This adaptation enables the generation of high-precision, spatially consistent change feature maps from bitemporal images without any supervision. Second, we design the Temporal Convolutional Gated Module (TCGM) within the STCU framework. The TCGM explicitly models spatio-temporal dependencies between dual-temporal features through multi-scale depthwise convolution and a gating mechanism, allowing it to capture both gradual transitions and abrupt changes while dynamically suppressing temporal noise. Furthermore, a multi-task loss function is devised to enhance feature disentanglement and joint optimization. Extensive experiments on three public benchmarks (CLCD, JL1, and RSIPAC) demonstrate that STCU significantly outperforms state-of-the-art unsupervised methods (e.g., SCM, FCD-GAN, Trans-MAD), achieving superior F1 scores of 53.69%, 60.37%, and 46.38%, respectively. The results confirm STCU's strong generalization capability and robustness in complex farmland environments, offering a practical and effective solution for large-scale agricultural monitoring.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: