Scholarlink

<p>Land Surface Temperature (LST) is an essential variable for environmental monitoring, yet obtaining high spatiotemporal resolution LST data is challenging due to the trade-off between spatial and temporal resolutions in thermal infrared sensors. Multiple algorithms have been developed for LST downscaling, including simple regression, data fusion, machine learning (ML), and hybrid methods. Among these, ML approaches offer a promising solution, but most studies focus on daytime LST and train models at a coarse spatial scale, assuming scale-invariant relationships with auxiliary data and often not incorporating fine-scale LST in model training. This limits their ability to capture extreme values. We downscaled both day and night Moderate Resolution Imaging Spectroradiometer (MODIS) LST observations from 1000&#xa0;m to 100&#xa0;m over India using fine-scale LST from either from the Landsat satellite or the Ecosystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) sensor as training data. Results show that models trained on fine-scale LST images acquired closer to MODIS overpass time were the most effective. Landsat-trained models worked well for daytime but struggled at night, requiring separate models using ECOSTRESS LST as a reference. Five ML models Random Forest, XGBoost, Artificial Neural Networks, ResNet18, and ResNet50 were evaluated using in-situ observations from 11 ground sites. Models trained with temporally matched fine-scale LST achieved the best performance, yielding RMSEs of 2.20–2.47&#xa0;K during the day and 1.97–2.41&#xa0;K at night for LST downscaling.</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

Machine Learning Based MODIS Day-Night Land Surface Temperature Downscaling Over India

  • 期刊论文

  • 发表日期：

    2026年4月4日

查看全文
  • Rahul Harod,
  • Eswar Rajasekaran,
  • Rahul Nigam,
  • Bimal Kumar Bhattacharya

摘要

Land Surface Temperature (LST) is an essential variable for environmental monitoring, yet obtaining high spatiotemporal resolution LST data is challenging due to the trade-off between spatial and temporal resolutions in thermal infrared sensors. Multiple algorithms have been developed for LST downscaling, including simple regression, data fusion, machine learning (ML), and hybrid methods. Among these, ML approaches offer a promising solution, but most studies focus on daytime LST and train models at a coarse spatial scale, assuming scale-invariant relationships with auxiliary data and often not incorporating fine-scale LST in model training. This limits their ability to capture extreme values. We downscaled both day and night Moderate Resolution Imaging Spectroradiometer (MODIS) LST observations from 1000 m to 100 m over India using fine-scale LST from either from the Landsat satellite or the Ecosystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) sensor as training data. Results show that models trained on fine-scale LST images acquired closer to MODIS overpass time were the most effective. Landsat-trained models worked well for daytime but struggled at night, requiring separate models using ECOSTRESS LST as a reference. Five ML models Random Forest, XGBoost, Artificial Neural Networks, ResNet18, and ResNet50 were evaluated using in-situ observations from 11 ground sites. Models trained with temporally matched fine-scale LST achieved the best performance, yielding RMSEs of 2.20–2.47 K during the day and 1.97–2.41 K at night for LST downscaling.