A novel land surface temperature reconstruction method and its application for downscaling surface soil moisture with machine learning.

• A new hybrid Land Surface Temperature (LST) reconstruction method is developed. • Performance indicators of the developed method are R2 = 0.94 and RMSE = 1.84°K. • Gapless LST data is used to downscale Surface Soil Moisture with Random Forest. • SMAP SSM L4 downscaled to 1 km accuracy with mean R (0.82) and ubRMSE (0.06 cm3/cm3). • Using antecedent precipitation as variable increased training performance (R2) from 0.76 to 0.93. Downscaling of soil moisture data is important for high resolution hydrological modeling. Most downscaling studies in the literature have used spatially discontinuous land surface temperature (LST) maps as the main auxiliary parameter, which limits the creation of continuous soil moisture maps. The number of studies on soil moisture downscaling with machine learning that use gapless LST maps is limited. With this motivation, a hybrid reconstruction method has been proposed in this study to practically obtain continuous LST maps, which are then used to produce high resolution surface soil moisture (SSM) datasets. The proposed method is shown to have high mean performance with R2 and RMSE values of 0.94 and 1.84°K, respectively, for the period between 2019 and 2022. The developed reconstructed LST maps were then used to downscale original 9 km spatial resolution soil moisture datasets of SMAP L3 and SMAP L4 with Random Forest (RF) machine learning algorithm. The RF model were run with four different rainfall datasets, and the MSWEP rainfall dataset was found to produce the best results. The use of antecedent rainfall values as input variables in machine learning models has been shown to improve the performance of the models R2 0.76 to 0.93. The accuracy of the downscaled data was later evaluated for Western Anatolia Basins (WAB) in Türkiye with 31 in-situ stations. The downscaled SMAP L4 had good average statistical indicators R (0.815 ± 0.1), RMSE (0.09 ± 0.047 cm3/cm3), and ubRMSE (0.058 ± 0.025 cm3/cm3). Downscaled SMAP L3 was also validated with in-situ observations with satisfactory R (0.79 ± 0.074), RMSE (0.09 ± 0.043 cm3/cm3), and ubRMSE (0.06 ± 0.026 cm3/cm3) statistics. Furthermore, the performance of the downscaled SMAP L3 was also cross validated with SMAP + Sentinel 1 (L2) dataset between 2019 and 2022. The mean statistics of R (0.761 ± 0.11) and Root Mean Squared Difference (RMSD) (0.05 ± 0.014 cm3/cm3) between downscaled SMAP L3 and L2 data revealed that the new reconstruction method of LST used in the RF model for downscaling of soil moisture performed well to obtain high resolution soil moisture datasets. The proposed technique also overcame the difficulties associated with coastal regions where data was masked for quality considerations, by not only enhancing overall spatial resolution but also filling these data gaps and giving a complete SSM coverage. [ABSTRACT FROM AUTHOR]