A non-stationary multi-scale data fusion framework for soil moisture estimation

Soil moisture (SM) has been identified as a key climate variable governing hydrologic and atmospheric processes across multiple spatial scales at local, regional and global levels. The global burgeoning of SM datasets in the past decade holds a significant potential in improving our understanding of multi-scale SM dynamics. The primary issues that hinder the fusion of different SM platforms are 1) different spatial resolutions of the data platforms, 2) inherent spatial variability in SM caused due to atmospheric and surface controls and 3) measurement errors caused due to imperfect retrievals of remote sensing platforms.

We present a novel data fusion scheme which takes all the above three factors into account using a Bayesian spatial hierarchical model (SHM). An SHM enables coherent integration of data, science and uncertainties to make optimal predictions at unobserved locations. Unlike popular machine learning methods, an SHM has direct physical interpretation enabling rich scientific inference and increased potential for transferability to data-scarce regions. To account for the inherent spatial variability in SM, we adapt a non-stationary spatial model such that the variance/correlation of SM varies with changing surface heterogeneity. The applicability of the fusion scheme is demonstrated by combining point, airborne and satellite data for two diverse watersheds (Iowa and Winnipeg) across continental North America. We quantify the effects of controls such as rainfall, vegetation and soil texture on the mean and the variance/correlation of SM as well as on the measurement errors of remote sensing platforms (C and L-band sensors). We find that that soil texture and vegetation affect the spatial distribution of SM at different stages of the crop growing cycle. While soil texture affects the spatial distribution of SM at the beginning of the crop growing cycle, vegetation becomes the dominant driver in peak biomass conditions. The bias in C-band sensor retrievals is affected by both soil and vegetation while that of L-band sensor is driven primarily by soil. We demonstrate that the proposed data fusion scheme is adept at assimilating and predicting SM distribution across all three scales (RMSE < 0.06 v/v) while accounting for potential measurement errors caused due to imperfect retrievals.