Integrating morphometric and climatic variables into the mapping of land use and vegetation cover in the Caatinga biome

Abstract Recent landscape transformations resulting from the replacement of natural areas with anthropogenic uses have had an impact on the semi-arid region of Brazil. The objective of the research was to evaluate a land use and land cover classification model in the Caatinga biome, using machine learning techniques integrated with multiple environmental and climatic covariates. The area was segmented into six subunits, based on a combination of hydrographic regions and geomorphological domains. The random forest algorithm on the Google Earth Engine platform was used to process the supervised classification, considering predictor variables, including spectral bands and indices, tasseled cap transformation, image fractions, surface temperature, precipitation, morphometric variables, and geographic location. The results indicated that geomorphological compartmentalization increased separability between land use classes and bare soil. The model achieved an average global accuracy of 0.83 and a kappa index of 0.80. The forest vegetation and rivers, lakes, and ocean classes showed high precision, while salt marshes and herbaceous restinga indicated inconsistencies. The most relevant variables were spatial position, altitude, and climatic conditions. In this perspective, incorporating this set of elements into the machine-learning classifier proved to be an innovative and scientifically relevant approach.