Estimation of Spatially Distributed Evapotranspiration Using Remote Sensing and a Relevance Vector Machine

Abstract

With the development of surface energy balance analyses, remote sensing has become a spatially explicit and quantitative methodology for understanding evapotranspiration (ET), a critical requirement for water resources planning and management. Limited temporal resolution of satellite images and cloudy skies present major limitations that impede continuous estimates of ET. This study introduces a practical approach that overcomes (in part) the previous limitations by implementing machine learning techniques that are accurate and robust. The analysis was applied to the Canal B service area of the Delta Canal Company in central Utah using data from the 2009–2011 growing seasons. Actual ET was calculated by an algorithm using data from satellite images. A relevance vector machine (RVM), which is a sparse Bayesian regression, was used to build a spatial model for ET. The RVM was trained with a set of inputs consisting of vegetation indexes, crops, and weather data. ET estimated via the algorithm was used as an output. The developed RVM model provided an accurate estimation of spatial ET based on a Nash-Sutcliffe coefficient (