Scholarlink

<p>High-resolution satellite imagery plays a pivotal role in remote sensing applications, including land-use and land-cover (LULC) mapping, urban planning, and environmental monitoring. Nevertheless, the spatial resolution of freely available satellite datasets, such as Sentinel-2 (S2), remains insufficient for fine-scale analyses. To overcome this limitation, super-resolution (SR) techniques have emerged as a powerful deep learning-driven approach for enhancing spatial detail beyond native sensor capabilities. In this study, we evaluate the impact of SR-enhanced S2 imagery (SRDR3; 10-band, 1&#xa0;m) on LULC classification performance over a heterogeneous landscape in Manavgat, Türkiye. The spectral and structural fidelity of SR outputs was assessed using RMSE, PSNR, SSIM, scatter plots, and bandwise R² metrics based on comparisons with Google Earth reference samples. Results show that SRDR3 preserves spectral consistency effectively in homogeneous classes. Water surfaces exhibited very low error (RMSE = 0.0075, PSNR = 29.13, SSIM = 0.9713, R² = 0.95), while agricultural areas remained within acceptable limits (RMSE = 0.0251, PSNR = 25.83, SSIM = 0.9217, R² = 0.90). In contrast, heterogeneous classes such as artificial surfaces and tree crops demonstrated higher distortion levels (RMSE = 0.0566–0.075; PSNR = 20.87–17.48; SSIM = 0.7857–0.7196; R² = 0.61–0.68).These differences were reflected in classification results: S2 achieved an overall accuracy (OA) of 96.7%, whereas SR achieved 94.2%. Although SR enhances spatial detail and supports accurate mapping in homogeneous environments, spectral distortions in structurally complex areas reduce class separability. The findings suggest that SR is a valuable complementary enhancement strategy, but preserving spectral fidelity remains essential for robust LULC applications.</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

Evaluating the impact of super-resolution on land use and land cover classification using sentinel-2 imagery

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Fatih Fehmi Şi̇mşek,
  • Kaan Kalkan

摘要

High-resolution satellite imagery plays a pivotal role in remote sensing applications, including land-use and land-cover (LULC) mapping, urban planning, and environmental monitoring. Nevertheless, the spatial resolution of freely available satellite datasets, such as Sentinel-2 (S2), remains insufficient for fine-scale analyses. To overcome this limitation, super-resolution (SR) techniques have emerged as a powerful deep learning-driven approach for enhancing spatial detail beyond native sensor capabilities. In this study, we evaluate the impact of SR-enhanced S2 imagery (SRDR3; 10-band, 1 m) on LULC classification performance over a heterogeneous landscape in Manavgat, Türkiye. The spectral and structural fidelity of SR outputs was assessed using RMSE, PSNR, SSIM, scatter plots, and bandwise R² metrics based on comparisons with Google Earth reference samples. Results show that SRDR3 preserves spectral consistency effectively in homogeneous classes. Water surfaces exhibited very low error (RMSE = 0.0075, PSNR = 29.13, SSIM = 0.9713, R² = 0.95), while agricultural areas remained within acceptable limits (RMSE = 0.0251, PSNR = 25.83, SSIM = 0.9217, R² = 0.90). In contrast, heterogeneous classes such as artificial surfaces and tree crops demonstrated higher distortion levels (RMSE = 0.0566–0.075; PSNR = 20.87–17.48; SSIM = 0.7857–0.7196; R² = 0.61–0.68).These differences were reflected in classification results: S2 achieved an overall accuracy (OA) of 96.7%, whereas SR achieved 94.2%. Although SR enhances spatial detail and supports accurate mapping in homogeneous environments, spectral distortions in structurally complex areas reduce class separability. The findings suggest that SR is a valuable complementary enhancement strategy, but preserving spectral fidelity remains essential for robust LULC applications.