Spatial prediction of evapotranspiration in a tropical mosaic landscape using remote sensing and explainable machine learning

Abstract Context Tropical forest regions are being transformed to other land covers at a high rate, often resulting in landscapes with heterogeneous vegetation structures. In combination with environmental factors this may influence ecosystem services, including evapotranspiration (ET). Objectives We aimed to predict the spatial variability of ET using spaceborne observations in a heterogenous tropical landscape, with a particular focus on elucidating the importance of vegetation structural characteristics and their interactions with environmental factors. Methods The study region was located in northeastern Madagascar. Daily ET was retrieved from ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS). Meteorological, topographical, soil and vegetation structure predictor variables were procured from various open sources. We applied forward feature selection and target oriented cross validation to address potential spatial autocorrelation and used SHAP analyses for elucidating interactions. Results Random forest models achieved high accuracies in the spatial prediction of ET, with R 2 values between 0.7 and 0.9 across different days. The explainable machine learning method, SHAP, revealed that highest contribution was by meteorological variables, followed by vegetation structure, topography, and soil. Analysis of key interactions between variables highlighted the role of vegetation structure in driving ET under different rainfall and wind speed conditions. Conclusions We conclude that the spatial variability of ET in this tropical mosaic landscape can be explained by a combination of biophysical variables, with vegetation structure contributing significantly. The underlying relationships can be useful to understand and potentially steer the climate regulation function of human-modified landscapes.