Scholarlink

<p>Shadows in high-resolution remote sensing images (HR-RSIs), caused by objects such as clouds, buildings, and trees, can provide essential information about an object’s shape, size, and texture but often distort images, hindering accurate recognition and extraction. Therefore, we propose a Shadow Aware Contrastive Network (SACNet) for shadow detection (SD) of HR-RSIs, leveraging self-supervised learning to extract features from unlabeled images. The SACNet effectively captures shadow features from unlabeled data in the self-supervised pretraining stage and adaptively adjusts the shadow predictions in the fine-tuning supervised network with a few labeled data based on the weights obtained from the self-supervised pretraining. Furthermore, there are only a few publicly available datasets for HR-RSI SD. To address this issue, we create a new HR-RSI SD dataset (SAAD) from high-resolution aerial images. Quantitative evaluations on two datasets showed that the SACNet achieved competitive overall accuracy, with the IoU of 81.51% and 73.72%, the F<sub>1</sub>-score of 92.27% and 86.12%, and the BER of 5.02% and 8.15% on the AISD and SAAD datasets, respectively. Comparative studies also confirmed the superiority and feasibility of the proposed SACNet for SD tasks, even with a limited number of labeled images.</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

SACNet: A Novel Self-Supervised Learning Method for Shadow Detection from High-Resolution Remote Sensing Images

  • 期刊论文

  • 发表日期：

    2025年2月24日

查看全文
  • Dehai Chen,
  • Jian Kang,
  • Lanying Wang,
  • Yongtao Yu,
  • Weixun Zhou,
  • Haiyan Guan,
  • Mannan Karim

摘要

Shadows in high-resolution remote sensing images (HR-RSIs), caused by objects such as clouds, buildings, and trees, can provide essential information about an object’s shape, size, and texture but often distort images, hindering accurate recognition and extraction. Therefore, we propose a Shadow Aware Contrastive Network (SACNet) for shadow detection (SD) of HR-RSIs, leveraging self-supervised learning to extract features from unlabeled images. The SACNet effectively captures shadow features from unlabeled data in the self-supervised pretraining stage and adaptively adjusts the shadow predictions in the fine-tuning supervised network with a few labeled data based on the weights obtained from the self-supervised pretraining. Furthermore, there are only a few publicly available datasets for HR-RSI SD. To address this issue, we create a new HR-RSI SD dataset (SAAD) from high-resolution aerial images. Quantitative evaluations on two datasets showed that the SACNet achieved competitive overall accuracy, with the IoU of 81.51% and 73.72%, the F1-score of 92.27% and 86.12%, and the BER of 5.02% and 8.15% on the AISD and SAAD datasets, respectively. Comparative studies also confirmed the superiority and feasibility of the proposed SACNet for SD tasks, even with a limited number of labeled images.