Prediction of Land Surface Temperature under overcast skies using Data Fusion and Deep Learning approach

Land Surface Temperature (Ts) is an essential input to drive surface energy balance for modelling terrestrial ecosystem processes. It serves as a vital indicator of drought, global change, urban heat islands, public health, and most importantly to understand monsoonal water stress signatures. Thermal InfraRed (TIR) remote sensing is the only source to retrieve Ts. Retrieval of Ts in the tropics and sub-tropics is regularly hindered due to persistent overcast skies during monsoon, limiting the seamless Ts data availability. Present study demonstrates the potential of Artificial Neural Network (ANN) to predict Ts over Indian landmass in overcast conditions using a three-stage multisource data-fusion approach. The Lout flux was predicted in the first stage using two five-layer deep learning-based ANN algorithms between 2015-2020. The trained ANN models provided a correlation (r) of 0.99 for daytime and 0.97 for night-time with Root Mean Square Error (RMSE) within 2%. The models’ outputs were upscaled to a spatial resolution of 4 km and compared with NCEP reanalysis fluxes (day and night-time algorithms combined) producing r = 0.83 and RMSE of 10%. Furthermore, Ts was retrieved using MODIS and other products and compared with ECMWF Reanalysis v5 (ERA5). The predicted LST provided a strong correlation (r = 0.98) with in-situ measurements leading to mean absolute error (MAE) of 2.4 K.