A Hybrid Physics-Based Method for Estimating Land Surface Temperature Using Radiative Transfer Simulations and Machine Learning Model from Sentinel-3A SLSTR Observations

Land surface temperature (LST) is crucial parameter in urban heat island studies, agricultural water management and drought monitoring. Study introduces a novel and efficient approach for retrieving LST from sea and land surface temperature radiometer onboard Sentinel-3A. A hybrid method has been developed using radiative transfer (RT) simulations, an explicit emissivity approach and advanced random forest machine learning (RF-ML) algorithm. The amalgamation of RT-modelling and ML offers a significant advantage in LST retrieval, where ML learns complex relationships directly from RT simulations bypassing the complex task of fitting numerous interrelated parameters. A total of 3,035,259 RT simulations encompassing globally representative conditions were generated with the MODTRAN 5.3 RT-model and used to train and validate the RF-ML model. The RF-ML model is demonstrably accurate, yielding less than 0.64 K root mean square error (RMSE) over RT simulations. The robustness of the method was further assessed through sensitivity analysis, highlighting its ability to handle higher uncertainties in water vapor and surface emissivity. The accuracy of a retrieved LST was validated using in-situ LST observations and obtained an overall systematic RMSE of 1.47 K with bias of -0.02 K over a 457-day period across diverse environments. These results demonstrate the efficacy of the proposed method in achieving accurate LST retrieval under diverse atmospheric and surface conditions. Furthermore, the method’s robustness in handling uncertainties associated with real-world applications, thereby making it a promising alternative for LST retrieval.