Avaliação da acurácia do mapeamento temático considerando imagens multiresolução

Abstract

The constant evolution of humanity has caused significant changes in the landscape, demanding efficient methods for monitoring land use and land cover changes. In this context, Remote Sensing stands out as a crucial tool for tracking and mitigating these changes, as well as supporting the planning of future interventions. Thus, this study evaluated the accuracy of thematic mapping of land use and land cover using images from the CBERS 04A (China–Brazil Earth Resources Satellite), Sentinel-2, and PRISMA (Hyperspectral Precursor of the Application Mission) satellites, aiming to identify the most efficient satellite. To conduct this study, images from the aforementioned satellites were processed using QGIS and ENVI Classic software, employing the supervised classification method to generate land use and land cover maps. During the classification process, the following algorithms were tested: Parallelepiped, Minimum Distance, Mahalanobis Distance, Maximum Likelihood, Neural Networks, and Support Vector Machine, to analyse the effectiveness of the classifiers in discriminating targets. The results were validated using confusion matrices, Kappa Index, and Overall Accuracy calculations. The results showed that the PRISMA satellite with the Mahalanobis Distance algorithm achieved the best performance among all classifications, with a Kappa Index of 0.9516 and an Overall Accuracy of 95.85%. The Sentinel-2 satellite showed consistent results across multiple classifiers, standing out as the most balanced option due to its excellent performance in most tested algorithms. The CBERS 04A yielded inferior but still robust results. Finally, land use and land cover maps were generated for the best classification models of each satellite.