A method for land surface temperature retrieval based on model-data-knowledge-driven and deep learning

Abstract

Most algorithms for land surface temperature (LST) retrieval depend on acquiring prior knowledge. To overcome this drawback, we propose a novel LST retrieval method based on model-data-knowledge-driven and deep learning, called the MDK-DL method. Based on the expert knowledge and radiation transfer model, we deduce LST retrieval mechanism and determine the best combination of the thermal infrared (TIR) bands of the sensor. Then, we use the radiation transfer model simulation and reliable satellite-ground data to establish a training and test database, and finally use the deep learning neural network for optimal computation. Three typical high-, medium- and low-spatial-resolution TIR remote sensing datasets (from Gaofen, the Moderate Resolution Imaging Spectroradiometer (MODIS), and Fengyun) are used for theoretical simulation and application analysis. The simulation shows that the minimum mean absolute error (MAE) is less than 0.1 K (standard deviation: 0.04 K; correlation coefficient: 1.000) at a small viewing direction (<7.5°) and less than 0.8 K at a large viewing direction (<65°). The in situ validation shows that the minimum MAE obtained by the optimal band combination is approximately 1 K (root mean square error (RMSE) = 1.12 K; coefficient of determination (R2) = 0.902). The retrieval accuracy is improved by increasing the number of TIR bands in the atmospheric window, and adding accurate atmospheric water vapor information produces better results. In general, four TIR bands in the atmospheric window bands are sufficient to retrieve the LST with high accuracy. Likewise, three TIR bands plus atmospheric water vapor information are sufficient for the retrieval requirements. All analyses indicate that our method is feasible and reliably accurate and can also be used to help design the instrument band to retrieve the LST with high precision.