Guan et al. (2026) A framework of coupling split-window and machine learning (SW-ML) for land surface temperature retrieval from MODIS thermal infrared data

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-04-08
• Authors: Yongjuan Guan, Si-Bo Duan, Songchao Chen, Xiaoxiao Min, Zhao-Liang Li
• DOI: 10.1016/j.rse.2026.115412

Research Groups

• State Key Laboratory of Efficient Utilization of Arable Land in China/Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China
• College of Environmental and Resource Sciences, Zhejiang University, Hangzhou 310058, China

Short Summary

This study proposes an SW-ML framework integrating physics-based split-window (SW) and machine learning (XGBoost) to accurately retrieve land surface temperature (LST) from MODIS thermal infrared data, demonstrating improved accuracy and interpretability, especially under challenging environmental conditions.

Objective

• To address limitations in existing LST retrieval methods by developing a robust, accurate, and interpretable SW-ML framework that integrates physics-based split-window (SW) and data-driven machine learning (ML) approaches to correct biases in SW-retrieved LST.

Study Configuration

• Spatial Scale: Global (based on the use of global in situ measurements for validation).
• Temporal Scale: Not explicitly defined as a fixed period, but the framework is designed for LST retrieval at specific observation times, with validation considering temporal extrapolation.

Methodology and Data

• Models used: Split-window (SW) algorithm, eXtreme Gradient Boosting (XGBoost), SW-ML framework (coupling SW and an XGBoost-based residual model).
• Data sources: MODIS thermal infrared data, high-quality global in situ LST measurements.

Main Results

• The SW-ML framework achieved a temporal Root Mean Square Error (RMSE) of 1.80 K and a spatial RMSE of 2.36 K during spatiotemporally independent validation, demonstrating good extrapolation capabilities.
• It consistently improved LST retrieval across most data partitions in ten-fold cross-validation, confirming its robustness.
• The framework outperformed the standalone SW algorithm by 1.0 K under high water vapor conditions, 0.49 K under elevated LST, and 1.2 K over barren land cover.
• SHAP analysis identified LST, water vapor, view time, elevation, longitude, and latitude as the six dominant factors influencing prediction residuals.
• LST exhibited a negative impact on predicted residuals, particularly at temperature extremes, while water vapor, view angle, and band 31 emissivity showed positive effects, leading to larger residuals under their high values.
• High accuracy was maintained even when trained on a very limited dataset (20% for training), significantly outperforming standalone ML methods.

Contributions

• Proposes a novel SW-ML framework that effectively integrates physics-based and data-driven approaches for robust and accurate LST retrieval.
• Achieves improved spatiotemporal extrapolation capabilities and consistent performance across diverse conditions.
• Provides interpretable predictions through SHAP analysis, identifying key driving factors for LST residuals.
• Enables effective bias correction of the traditional SW algorithm, particularly under challenging environmental conditions.
• Demonstrates high accuracy even with limited training data, enhancing practical applicability.

Funding

• Not mentioned in the provided paper text.

Citation

@article{Guan2026framework,
  author = {Guan, Yongjuan and Duan, Si-Bo and Chen, Songchao and Min, Xiaoxiao and Li, Zhao-Liang},
  title = {A framework of coupling split-window and machine learning (SW-ML) for land surface temperature retrieval from MODIS thermal infrared data},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115412},
  url = {https://doi.org/10.1016/j.rse.2026.115412}
}