Aprendizagem de máquina para detecção de áreas de aluvião: um estudo no semiárido brasileiro

Abstract

In semi-arid regions, where water resources are scarce, the development of precise methods for their identification and mapping is pertinent. This study aimed to investigate the use of artificial intelligence techniques to map alluvial deposits in the Riacho Grande basin, Pernambuco. To this end, an extensive, high-resolution, georeferenced database was employed, comprising eleven geomorphological and hydrological dimensions for the characterisation of alluvial areas. Initially, the database, containing over 315 million points distributed across the Basin, underwent an optimisation process using the extended Geometric Markovian Diffusion (e-GMD) method, with the aim of reducing the data volume at different levels without compromising its representativeness. Subsequently, supervised machine learning models (KNN, CART/DT, and RF) were constructed and evaluated for the detection of alluvial areas. Each model was trained and validated using 10-fold cross-validation, and its hyperparameters were tuned to optimise performance. The results demonstrated that the Random Forest (RF) algorithm excelled, achieving an F1-score of 89.8% using only 5% of the original data, outperforming KNN (86.1%) and CART/DT (86.3%). This study demonstrates the potential of employing artificial intelligence techniques for the effective mapping of alluvial deposits in semi-arid regions.