Scholarlink

<p>To address the spectral confusion caused by similar crop growth stages and seasonal crop rotation, existing remote sensing change detection methods have significant limitations in distinguishing between short-term crop rotation and long-term non-food crop transitions in farmland. These limitations primarily manifest in limited ability to model long-range temporal dependencies, imperfect multi-source information fusion mechanisms, and a lack of effective discrimination against the persistence of changes. Therefore, this study proposes a remote sensing change detection framework that integrates a Transformer model and spatiotemporal features to identify persistent non-food crop transitions in farmland. This framework employs a hierarchical spatial encoder to extract multi-scale spatial representations, combines a temporal Transformer with time-interval location coding to model long-range temporal dependencies, and adaptively fuses optical, SAR, and seasonal features through a cross-attention mechanism. To further enhance the ability to distinguish between short-term disturbances and persistent changes, a temporal consistency loss and a two-branch discriminant head are introduced. Experimental results show that the proposed method achieves pixel-level F1 scores of 0.87 and IoU of 0.78 on multi-regional field samples and public benchmark datasets, with a recall rate of 0.91 for non-food crops, a misclassification rate of only 11.0% for short-term crop rotation, an average detection latency of 12&#xa0;days, and an expected calibration error of 0.06, significantly outperforming existing state-of-the-art methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic monitoring of farmland non-grain conversion based on transformer and multi-source spatiotemporal features

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Xianguang Ma,
  • Feng He,
  • Bohui Tang

摘要

To address the spectral confusion caused by similar crop growth stages and seasonal crop rotation, existing remote sensing change detection methods have significant limitations in distinguishing between short-term crop rotation and long-term non-food crop transitions in farmland. These limitations primarily manifest in limited ability to model long-range temporal dependencies, imperfect multi-source information fusion mechanisms, and a lack of effective discrimination against the persistence of changes. Therefore, this study proposes a remote sensing change detection framework that integrates a Transformer model and spatiotemporal features to identify persistent non-food crop transitions in farmland. This framework employs a hierarchical spatial encoder to extract multi-scale spatial representations, combines a temporal Transformer with time-interval location coding to model long-range temporal dependencies, and adaptively fuses optical, SAR, and seasonal features through a cross-attention mechanism. To further enhance the ability to distinguish between short-term disturbances and persistent changes, a temporal consistency loss and a two-branch discriminant head are introduced. Experimental results show that the proposed method achieves pixel-level F1 scores of 0.87 and IoU of 0.78 on multi-regional field samples and public benchmark datasets, with a recall rate of 0.91 for non-food crops, a misclassification rate of only 11.0% for short-term crop rotation, an average detection latency of 12 days, and an expected calibration error of 0.06, significantly outperforming existing state-of-the-art methods.