Mapeamento da cobertura e uso da terra e da vegetação arbórea utilizando dados espectrais e topográficos do PE3D na região metropolitana do Recife-PE

Abstract

Land cover and land use mapping, as well as urban tree vegetation mapping, are essential for supporting territorial planning, environmental monitoring, and urban management. However, the high complexity of urban landscapes and the spectral similarity among different land cover classes make remote sensing image classification a challenging task, requiring methods capable of integrating multiple sources of information. In this context, this study evaluated the performance of machine learning algorithms for land cover and land use mapping using data from the Pernambuco Tridimensional Project (PE3D). Spectral (RGB and IHS) and topographic (DTM, nDSM, and slope) attributes derived from very high spatial resolution imagery were employed in the region encompassing the municipalities of Recife, Camaragibe, Olinda, Paulista, and São Lourenço da Mata, Pernambuco, Brazil. Supervised classifications were performed i n QGIS using the Dzetsaka plugin and the Gaussian Mixture Model (GMM), Random Forest (RF), and Extra Trees (ET) algorithms. Different classification scenarios were compared to evaluate the contribution of spectral and topographic attributes to model performance. In addition to the multiclass land cover and land use classification, a binary classification of Tall Vegetation and Non-Tall Vegetation was carried out to identify urban trees and remnants of urban forests. The results showed that the integration of RGB, IHS, DTM, nDSM, and slope attributes provided the best classification performance. The Extra Trees algorithm achieved the highest overall accuracy (86.1%), followed by Random Forest (84.9%) and Gaussian Mixture Model (80.6%). For the Tall Vegetation mapping, overall accuracies above 97% were achieved, demonstrating the potential of very high spatial resolution imagery for urban tree canopy identification. Furthermore, decision tree-based algorithms showed a greater ability to discriminate classes with similar spectral and topographic characteristics. It is therefore concluded that the integration of spectral and topographic attributes significantly improves land cover and land use mapping, as well as the identification of urban tree vegetation, highlighting the potential of PE3D data and machine learning algorithms for remote sensing, territorial planning, and environmental monitoring applications.