Spatial Root Zone Soil Moisture Estimation and Forecasting Using the METRIC Evapotranspiration Product and Multivariate Relevance Vector Machines

Document Type


Journal/Book Title/Conference

American Geophysical Union Conference, San Francisco, CA, USA

Publication Date



Limited access to spatial root zone soil moisture (SM) estimation in agricultural areas restricts enhanced water balance and irrigation scheduling estimations by irrigators and water managers, as well as other possible uses of these soil moisture estimates. Herein, we propose a methodology that allows for spatial SM estimation and forecasts at depths of 0.05, 0.30 and 0.60 m in agricultural areas at a temporal resolution ranging from the present to eight and sixteen days ahead. This methodology is based on a statistical learning model called the Multivariate Relevance Vector Machine (MVRVM). This model is known for its robustness, efficiency, and sparseness. It provides a statistically sound approach to learn from the input-output response patterns contained in the training dataset, and has proven to be superior to traditional algorithms such as Artificial Neural Networks. The MVRVM is used to build a methodology that spatially estimates and predicts current and future soil moisture state based upon historical records of soil moisture and actual crop evapotranspiration. Soil moisture measurements at three different depths acquired by the Utah Water Research Laboratory (UWRL) for agricultural lands in the Lower Sevier River Basin, Utah, are used for this study. The methodology combines the SM data at different depths along with estimates of actual crop evapotranspiration using the Mapping EvapoTranspiration at high Resolution with Internalized Calibration (METRIC) algorithm which uses Landsat TM and ETM+ imagery records. The MVRVM produces good results at current, eight and sixteen days with a reduced computational complexity and suitable real-time implementation. Additionally, spatial bootstrapping analysis is used to evaluate over- and under-fitting and uncertainty in model estimates.

This document is currently not available here.