GROUNDWATER RECHARGE AREAS MAPPING IN HIGH-ALTITUDE ANDEAN MOUNTAINS THROUGH MACHINE LEARNING ALGORITHMS

Abstract

The high-altitude wetlands in the Central Andes are unique ecosystems located above 4000 masl in the Bolivian Altiplano. The analysis and classification of spatial information is a crucial step in the identification of wetlands in scarped topography. The objective of this study was to test machine learning algorithms to map Andean wetlands. The first step consisted on applying the machine learning algorithms Least Absolute Shrinkage and Selection Operator LASSO and Receiver Operating Characteristic ROC for the sensitivity analysis. Then, there were compared the Random Forest Regressor RFR, Support Vector Regressor SVR, and Multivariate Adaptative Regression Splines MARS regression supervised machine learning algorithms for the wetlands mapping. Results were validated by Google Earth satellite images and a regression coefficient. The RFR showed good results for areas with slopes of 0 - 32 degrees; the SVR showed good performance for areas with slopes of 44 - 76 degrees, while for areas with slopes of 0 - 12 degrees its performance was inaccurate. The application of the MARS showed trivial results compared to those obtained by the first two algorithms; some results were good for certain areas, areas with slopes of 0 - 12 degrees and 44 - 77 degrees were erroneously flagged.