Improved fusion model for generating hourly fine scale land surface temperature data under all-weather condition

Abstract

Existing Land Surface Temperature (LST) fusion models encounter some challenges due to missing data, complex weather areas, and rapid land cover changes. To overcome these limitations, we proposed the Integrated SpatioTemporal Fusion Algorithm (ISFAT). ISFAT is developed based on contemporary fusion models but in addition incorporates data from partially contaminated LSTs using the masked weight function. This helps to predict fine-scale LST on prediction date while considering error resulting from landcover changes between the base and prediction date. This algorithm also factors in the calculation of model residuals, which are distributed back to the predicted fine-scale LST using the thin-plate spline function. The fine-scale LST on prediction can thereafter employed for predicting hourly fine-scale LST images by integrating a coarse resolution LST with hourly temporal resolution. Compared to contemporary LST fusion models, ISFAT demonstrates superior performance, with mean average differences of 0.1 K and 0.27 K over SADFAT and STITFM, respectively. Additionally, diurnal LST predictions from ISFAT compare well with air temperatures from automatic weather stations. Notably, on February 18, 2020, ISFAT effectively optimized fine-scale LST for Hong Kong, achieving an RMSE of 3.33 K, despite the limitation of cloud cover in the base date. The newly developed ISFAT could facilitate better LST retrieval over a large spatial coverage under different degrees of cloud contamination.