GEOPROCESSING AND CONVOLUTIONAL NEURAL NETWORKS: ANALYSIS OF LAND COVER AND LAND USE IN THE ALMADA RIVER BASIN (STATE OF BAHIA, BRAZIL)

Abstract

Geoprocessing techniques associated with convolutional neural network models (CNN) emerge as a viable alternative for obtaining data to subsidize decision-making. In this context, this work aimed to evaluate the application of CNN algorithms for the classification and detection of land cover and use classes in satellite images of the Almada River Basin (ARB). To achieve the goal, logical steps were structured: (i) information collection; (ii) processing of Dataset Eurosat; and (III) the evaluation of the models. Classification results demonstrated more than 90% precisions in class recognition. As for the detection model, a 70% accuracy was identified for the “Forest” and “Pasture” classes, which have large extensions within ARB. Both models showed their versatility in application and viability as tools for monitoring the physical and environmental conditions of ARB. In this sense, the effectiveness of the models is emphasized in the identification and location of land cover and use classes, emphasizing the importance of building a dataset that highlights the characteristics of the study area. This contributes to obtaining reliable results and improving the practical use of CNN models.