MULTICRITERIA ANALYSIS AND MACHINE LEARNING ALGORITHM FOR DEFINITION OF AREAS FOR MICRO-DAM, SOUTHEASTERN BRAZIL

Abstract

Micro-dams are efficient in controlling erosion and increases the rate of water infiltration into the soil. The objective was to determine potential areas for its construction. We used a digital elevation model (DEM) data and land use map. We prepared the map of potential areas for micro-dam from the multicriteria analysis (AHP). On the map, we determined a grid of 2000 points for extracting values. Also, we developed a methodological framework to predict data (points) with the Random Forest (RF) algorithm, with the selection of relevant covariates from DEM.  The final map results from the union of the maps of the two techniques (AHP+RF). The AHP method overestimates high potential areas in zones without potential. The RF model used seven topographic covariates, and they are related to hydrological flows. The R² performance was 0.43; however, the RF model was not efficient in determining areas of low potential. The isolation of the low potential areas of the AHP method overlapping on the RF map generated a more consistent map, favoring a 99% reduction in areas of very high potential in areas with no aptitude. The medium potential predominates in the study area (44%), and areas of high potential occupy 5.15%.