Soil moisture (SM) at the interface between the land surface and atmosphere is one of the major environmental parameters which plays an important role in hydrological applications. In this article, the SM measured by Soil Moisture Active Passive (SMAP) is downscaled from 3- to 1-km spatial resolution. The main purpose is to evaluate the performance of two different downscaling methods over a variety of climatic conditions and land cover types. These two methods, based on regression and artificial neural network (ANN), are used for enabling us to cross-validate and reliably interpret the obtained results since the number of ground measurements is not sufficient for accuracy assessment. These methods are applied over four different case studies; one is located in the USA, i.e., state of Utah (semi-arid), and the remaining three are located in Iran, i.e., Fars (arid and semi-arid), Yazd (hyper-arid), and Golestan (humid). In both methods, different combinations of input features correlated with SM including land surface temperature (LST), normalized difference vegetation index (NDVI), brightness temperatures in horizontal and vertical polarizations (TBH and TBV), shortwave infrared (SWIR), and digital elevation model (DEM) are used. It is found that the DEM does not add extra information in downscaling. The reason is due to high correlation between topography and LST. Moreover, SWIR is most likely able to model only large-scale variations of SM. The downscaled SM products are then compared to 1-km resolution SMAP SM extracted from Sentinel-1 for the study areas in Iran and in situ measurements in Utah. Both methods produce results which are considerably consistent except that the regression method adds more spatial details in the downscaled SM. The achievements illustrate that the performance of both downscaling methods is higher in areas with more homogeneous climatic conditions, i.e., Yazd and Golestan. The best evaluation metrics including correlation coefficient (R), root mean square error (RMSE), and mean absolute percentage error (MAPE) for Yazd and Golestan are R = 0.89, RMSE = 0.025 m3/m3, and MAPE = 21.13% and R = 0.93, RMSE = 0.044 m3/m3, and MAPE = 21.95%, respectively. Moreover, large model biases are associated with dense vegetated areas and high altitudes. The best downscaling accuracy in both methods over all study areas belongs to bare soil and flat regions.
Keywords: Downscaling; Neural network; Regression; SMAP; Soil moisture.
© 2023. The Author(s), under exclusive licence to Springer Nature Switzerland AG.