Importance of vegetation structure for predicting evapotranspiration in a tropical mosaic landscape

Abstract

Evapotranspiration (ET) is a critical process within the hydrological cycle, susceptible to shifts due to changes in land use. In tropical forest regions, widespread transformations often result in mosaic patterns of land-use types. Our goal was to explore the importance of vegetation structure, topography, meteorology and soil for the spatial variability of ET in a tropical mosaic landscape. We used a random forest machine learning technique for spatial data, employing forward feature selection and cross-validation to prevent overfitting. Our study region is situated in north-eastern Madagascar and is mainly composed of forest fragments, vanilla agroforests, rice fields and fallow land of shifting cultivation. We used a combination of open-source data products derived from various satellite experiments. Daily ET data were retrieved from the ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS). Forest structure predictors from GEDI and PROBA-V, meteorological data from ERA5, topography from JAXA and soil data from ISRIC were obtained. The variables included in the L3 algorithm to calculate ECOSTRESS ET daily data were not included in the study to prevent bias in the models. The models achieved high accuracy for the spatial prediction of ET (R2) of 0.76 and 0.82 for different days. Besides other biophysical variables, leaf area index, tree cover and tree height were important variables in predicting ET. Our findings thereby underscore the crucial role of forest structure on ET even in complex structured tropical mosaic landscapes.