Wavelet-multivariate relevance vector machine hybrid model for forecasting daily evapotranspiration

Evapotranspiration (ET) is one of the main components of the hydrological cycle. It is a complex process driven mainly by weather parameters, and as such, is characterized by high non-linearity and non-stationarity. This paper introduces a methodology combining wavelet multiresolution analysis with a machine learning algorithm, the multivariate relevance vector machine (MVRVM), in order to predict 16 days of future daily reference evapotranspiration (ETo). This methodology lays the ground for forecasting the spatial distribution of ET using Landsat satellite imagery, hence the choice of 16 days, which corresponds with the Landsat overpass cycle. An accurate prediction of daily ETo is needed to improve the management of irrigation schedules as well as the operations of water supply facilities like canals and reservoirs. In this paper, various wavelet decompositions were performed and combined with MVRVM to develop hybrid models to predict ETo over a 16-days period. These models were compared to a MVRVM model, and models accuracy and robustness were evaluated. The addition of 10 days of forecasted air temperature as additional inputs to the forecasting models was also investigated. The results of the wavelet-MVRVM hybrid modeling methodology showed that a reliable forecast of ETo up to 16 days ahead is possible.