Downscaling and forecasting of evapotranspiration using a synthetic model of wavelets and support vector machines

Providing reliable forecasts of evapotranspiration (ET) at farm level is a key element toward efficient water management in irrigated basins. This paper presents an algorithm that provides a means to downscale and forecast dependent variables such as ET images. Using the discrete wavelet transform (DWT) and support vector machines (SVMs), the algorithm finds multiple relationships between inputs and outputs at all different spatial scales and uses these relationships to predict the output at the finest resolution. Decomposing and reconstructing processes are done by using 2-D DNVT with basis functions that suit the physics of the property in question. Two-dimensional DWT for one level will result in one datum image (low-low-pass filter image) and three detail images (low-high, high-low, and high-high). The underlying relationship between the input variables and the output are learned by training an SVM on the datum images at the resolution of the output. The SVM is then applied on the detailed images to produce the detailed images of the output, which are needed to help downscale the output image to a higher resolution. In addition to being downscaled, the output image can be shifted ahead in time, providing a means for the algorithm to be used for forecasting. The algorithm has been applied on two case studies, one in Bondville, IL, where the results have been validated against AmeriFlux observations, and mother in the Sevier River Basin, UT.