Fusion of remotely sensed data for soil moisture estimation using relevance vector and support vector machines

Bushra Zaman, Utah State University
Mac McKee, Utah State University
Christopher M. U. Neale, Utah State University

Abstract

A data assimilation (DA) methodology that uses two state-of-the-art techniques, relevance vector machines (RVMs) and support vector machines (SVMs), is applied to retrieve surface (0–6 cm) soil moisture content (SMC) and SMC at a depth of 30 cm. RVMs and SVMs are known for their robustness, efficiency and sparseness and provide a statistically sound approach to solve inverse problems and thus to build statistical models. Here, we build a statistical model that produces acceptable estimations of SMC by using inexpensive and readily available data. The study area for this research is the Walnut Creek watershed in Ames, south-central Iowa, USA. The data were obtained from Soil Moisture Experiments 2002 (SMEX02) conducted at Ames, Iowa. The DA methodology combines remotely sensed inputs with field measurements, crop physiological characteristics, soil temperature, soil water-holding capacity and meteorological data to build a two-step model to estimate SMC using both techniques, i.e. RVMs and SVMs. First, the RVM is used to build a model that retrieves surface (0–6 cm) SMC. This information serves as a boundary condition for the second step of this model, which estimates SMC at a depth of 30 cm. An exactly similar routine is followed with an SVM for estimation of surface (0–6 cm) SMC and SMC at a depth of 30 cm. The results from the RVM and SVM models are compared and statistics show that RVMs perform better (root mean square error (RMSE) = 0.014 m3 m-3) when compared with SVMs (RMSE = 0.017 m3 m-3) with a reduced computational complexity and more suitable real-time implementation. Cross-validation techniques are used to optimize the model. Bootstrapping is used to check over/under-fitting and uncertainty in model estimates. Computations show good agreement with the actual SMC measurements with coefficients of determination (R2) for RVM equal to 0.92 and for SVM equal to 0.88. Statistics indicate a good model generalization capability with indexes of agreement (IoAs) for RVM equal to 0.97 and for SVM equal to 0.96.