Scholarlink

<p>Obtaining high spatiotemporal resolution Land Surface Temperature (LST) data (≤ 100&#xa0;m and ~daily) is pivotal for environmental studies, yet it has been challengeing due to the spatiotemporal trade-offs of remote sensing data, further complicated by cloud-cover-induced data gaps. Previous methods have typically focused on either spatial downscaling or on gap-filling techniques individually, leaving the integrative benefits of these techniques largely unexplored. Additionally, machine learning models for LST prediction often rely solely on linear or non-linear regression, overlooking the benefits of ensemble stacking for improving accuracy. Here, we present a hybrid model that integrates spatial reconstruction with kernel-based downscaling through ensemble stacking, utilizing Extra Trees (ET) and Ridge regressions. This approach leverages variables including surface reflectance, albedo, spectral indices, digital elevation model, and land cover data to generate seamless 100&#xa0;m LST data from the 1 km Moderate Resolution Imaging Spectroradiometer (MODIS) imageries with less than 40% cloud cover across nine Australian OzFlux sites from 2013 to 2021. Through in-situ validations of reconstructed LST patches, the hybrid model demonstrated better accuracy, with a Pearson’s correlation coefficient (<i>R</i>) of 0.95, a bias of 1.45&#xa0;K, a Root Mean Square Error (RMSE) of 3.55&#xa0;K, and a Mean Absolute Error (MAE) of 2.71&#xa0;K compared to both ET (<i>R</i> = 0.95, bias = 1.5&#xa0;K, RMSE = 3.57&#xa0;K, MAE = 2.72&#xa0;K) and Ridge (<i>R</i> = 0.94, bias = 1.43&#xa0;K, RMSE = 3.62&#xa0;K, MAE = 2.74&#xa0;K). By further evaluating against the ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) LST data, the hybrid model showed improved correlation (<i>R</i> = 0.55) and the lowest error metrics (RMSE = 3.86&#xa0;K, MAE = 3.34&#xa0;K), outperforming both ET (<i>R</i> = 0.53) and Ridge (<i>R</i> = 0.44). The downscaled LST patches also maintained robust performance, illustrating minimal differences between the hybrid and individual models in both in-situ and ECOSTRESS comparisons. Comparative assessments of the downscaled LST over diverse climatic regions indicate that the hybrid model performed better than the established benchmark models.</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

Generation of 100 m seamless land surface temperature from clear sky or partially cloudy MODIS data using Landsat-assisted stacked ensemble regression

  • 期刊论文

  • 发表日期：

    2025年11月4日

查看全文
  • Zahid Jahangir,
  • Zhenfeng Shao,
  • Yi Yu,
  • Peng Fu,
  • Qazi Muhammad Yasir,
  • Xueying Xiao

摘要

Obtaining high spatiotemporal resolution Land Surface Temperature (LST) data (≤ 100 m and ~daily) is pivotal for environmental studies, yet it has been challengeing due to the spatiotemporal trade-offs of remote sensing data, further complicated by cloud-cover-induced data gaps. Previous methods have typically focused on either spatial downscaling or on gap-filling techniques individually, leaving the integrative benefits of these techniques largely unexplored. Additionally, machine learning models for LST prediction often rely solely on linear or non-linear regression, overlooking the benefits of ensemble stacking for improving accuracy. Here, we present a hybrid model that integrates spatial reconstruction with kernel-based downscaling through ensemble stacking, utilizing Extra Trees (ET) and Ridge regressions. This approach leverages variables including surface reflectance, albedo, spectral indices, digital elevation model, and land cover data to generate seamless 100 m LST data from the 1 km Moderate Resolution Imaging Spectroradiometer (MODIS) imageries with less than 40% cloud cover across nine Australian OzFlux sites from 2013 to 2021. Through in-situ validations of reconstructed LST patches, the hybrid model demonstrated better accuracy, with a Pearson’s correlation coefficient (R) of 0.95, a bias of 1.45 K, a Root Mean Square Error (RMSE) of 3.55 K, and a Mean Absolute Error (MAE) of 2.71 K compared to both ET (R = 0.95, bias = 1.5 K, RMSE = 3.57 K, MAE = 2.72 K) and Ridge (R = 0.94, bias = 1.43 K, RMSE = 3.62 K, MAE = 2.74 K). By further evaluating against the ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) LST data, the hybrid model showed improved correlation (R = 0.55) and the lowest error metrics (RMSE = 3.86 K, MAE = 3.34 K), outperforming both ET (R = 0.53) and Ridge (R = 0.44). The downscaled LST patches also maintained robust performance, illustrating minimal differences between the hybrid and individual models in both in-situ and ECOSTRESS comparisons. Comparative assessments of the downscaled LST over diverse climatic regions indicate that the hybrid model performed better than the established benchmark models.