Daily Evapotranspiration Mapping Using Regression Random Forest Models

Efficient water management in agriculture requires an accurate estimation of evapotranspiration (ET). Even though local measurements can be used to estimate the components of the surface energy balance, these values cannot be extrapolated to large areas due to the heterogeneity and complexity of agricultural and natural land surfaces and the dynamic nature of their heat processes. This extrapolation can be done by using satellite imagery, which provides information in the infrared thermal band; however, this band is not available in most current operational remote sensors. Our work hypothesis is that it is possible to generate a spatially distributed estimation of $\text{ET}_{d}$ without thermal band by using nonparametric models as regression random forest models (RRFM). Six Landsat-7 scenes were used to generate the RRFM. Results were evaluated by comparing the values of $\text{ET}_{d}$ provided by RRFM with that obtained using the surface energy balance model. It has been shown that the results generated by RRFM present a good agreement with METRIC (mapping ET at high resolution using internalized calibration) results, both quantitatively and qualitatively, especially for agricultural vegetation and forest land covers. Moreover, it has been detected that the RRFM estimation quality depends on the meteorological conditions on the days previous to the satellite register. It can be concluded that the $\text{ET}_{d}$ estimated by the RRFM would be feasible for real applications when the thermal band is not available.