An evaluation of nonlinear methods for estimating catchment-scale soil moisture patterns based on topographic attributes

Physical processes that impact soil moisture are typically expressed as nonlinear functions, but most previous research on the estimation of soil moisture has relied on linear techniques. In the present work, two machine learning techniques, a spatial artificial neural network (SANN) and a mixture model (MM), that can infer nonlinear relationships are compared to multiple linear regression (MLR) for estimating soil moisture patterns using topographic attributes as predictor variables. The methods are applied to time-domain reflectometry (TDR) soil moisture data collected at three catchments with varying characteristics (Tarrawarra, Satellite Station and Cache la Poudre) under different wetness conditions. The methods' performances with respect to the number of predictor attributes, the quantity of training data and the attributes employed are compared using the Nash Sutcliffe coefficient of efficiency (NSCE) as the performance measure. The performances of the methods are dependent on the site studied, the average soil moisture and the quantity of training data provided. Although the methods often perform similarly, the best performing method overall is the SANN, which incorporates additional predictor variables more effectively than the other methods.