Scholarlink

<p>Reliable land use and land cover (LULC) classification is critical for environmental observation and geospatial analysis from remote sensing images. Although transformer models have shown promising results, the reliability and predictive uncertainty of these models have not been adequately investigated in the context of remote sensing. This paper proposes a reliability-aware LULC classification framework with a Swin Transformer model on the EuroSAT RGB dataset. In addition to the accuracy assessment, the proposed framework also includes model calibration analysis, predictive uncertainty estimation, selective prediction, and out-of-distribution (OOD) detection for enhancing the reliability of predictions. The experimental outcomes show a classification accuracy of 98.70 with a remarkably low Expected Calibration Error of 0.0019 from 0.02, which reveals well-calibrated confidence estimates. The confidence-based selective prediction further enhances the reliability of predictions with 99.73 accuracy at 92.3 coverage, indicating enhanced reliability by rejecting uncertain predictions, while the OOD detection performance is 0.93 AUROC, exhibiting strong discrimination between in-distribution and out-of-distribution samples. The proposed framework provides accurate and reliable LULC classification under a benchmark setting, with well-calibrated and steady predictions.</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

A reliability-aware Swin Transformer framework for land use and land cover classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月20日

查看全文
  • Shankesh Raja,
  • R. Srinidhi,
  • R. Gayathri,
  • R. Dayana

摘要

Reliable land use and land cover (LULC) classification is critical for environmental observation and geospatial analysis from remote sensing images. Although transformer models have shown promising results, the reliability and predictive uncertainty of these models have not been adequately investigated in the context of remote sensing. This paper proposes a reliability-aware LULC classification framework with a Swin Transformer model on the EuroSAT RGB dataset. In addition to the accuracy assessment, the proposed framework also includes model calibration analysis, predictive uncertainty estimation, selective prediction, and out-of-distribution (OOD) detection for enhancing the reliability of predictions. The experimental outcomes show a classification accuracy of 98.70 with a remarkably low Expected Calibration Error of 0.0019 from 0.02, which reveals well-calibrated confidence estimates. The confidence-based selective prediction further enhances the reliability of predictions with 99.73 accuracy at 92.3 coverage, indicating enhanced reliability by rejecting uncertain predictions, while the OOD detection performance is 0.93 AUROC, exhibiting strong discrimination between in-distribution and out-of-distribution samples. The proposed framework provides accurate and reliable LULC classification under a benchmark setting, with well-calibrated and steady predictions.