Comparison of two satellite-based soil moisture reconstruction algorithms: A case study in the state of Oklahoma, USA

Abstract

Soil moisture plays a critical role in regional water cycles and is one of the key indicators for agricultural droughts. Satellite-based microwave radiometers have been the main instrument for mapping surface soil moisture at regional to global scales. However, known issues still exist that lead to data gaps in microwave-based remotely sensed soil moisture products. Multiple gap-filling approaches have been recently developed to generate spatially continuous satellite-based soil moisture. However, there are few inter-comparisons of such approaches despite their importance. In this study, we present a comparison between a triangular feature space-based (Tri) model and a machine learning (ML) based random forest (RF) model for seamless reconstruction in the European Space Agency’s Essential Climate Variables Soil Moisture product (ECV SM) over Oklahoma, USA. Five variables are implemented in the models: the normalized difference vegetation index (NDVI), daytime land surface temperature (LST), nighttime LST, daily average LST, and diurnal LST (ΔLST), and different combinations of these five variables are examined. The reconstructed soil moisture is validated against the original ECV SM and local in-situ measurements. The RF model achieved precise and outstanding performance (R2 = 0.95, RMSE = 0.02 m3/m3, bias = 0.07%) in comparison to the original ECV SM. Comparatively, the Tri and RF models revealed equivalent performance in fitting the in-situ measurements, and both effectively alleviated the bias from the ECV SM. Specifically, Tri, which used daytime LST and NDVI to establish a soil moisture estimation model, displayed the highest correlation coefficient as well as the smallest error (R = 0.620, RMSE = 0.080 m3/m3, ubRMSE = 0.038 m3/m3) for the in-situ measurements. Moreover, the assessments for sub-regions proved the consistency and robustness of both the Tri and RF models. Missing sub-regional training samples are unlikely to degrade performance of the models as long as there are sufficient valid samples in the whole region. The results of this study highlight the promising potential of Tri and RF models in efficiently filling the gaps in the ECV SM and can provide a reference for future studies focused on satellite-based soil moisture gap-filling algorithm selections.