Sensoriamento Remoto aplicado em Recursos Hídricos

Abstract

The monitoring of water resources serves as a basis for decision making and even to mitigate the effects of future water crises, such as the crisis in the Cantareira System, the study area of this work, in the 2013/14 biennium. We investigated the reliability of image classification, using remote sensing techniques and machine learning in the context of water resources, which is an indispensable resource for society. The experiments were carried out in the six dams that make up the Cantareira System, and RapidEye orbital multispectral images were used, which have a spatial resolution of 5 meters. Four classification methods were tested, namely: Minimum Distance, Maximum Likelihood, Spectral Angle Mapping and Random Forest. The Minimum Distance and Maximum Likelihood methods offered results with an accuracy greater than 95%. The Random Forest, a machine learning technique, made it possible to generate results with superior accuracy, reaching an accuracy of 98.06%. The results show that the combination of RapidEye images with remote sensing and machine learning techniques allows detailed and accurate mapping of water resources in the Cantareira System. As a result of this research, there is also the generation of a set of labeled data, available for future experiments.