Deep Learning para Classificação Supervisionada de Imagens CBERS-4A da Área Urbana de Rio Claro (SP)

This study evaluates the accuracy of land use and land cover (LULC) mapping in an urban area of Rio Claro (SP) using Deep Learning techniques and a CBERS-4A (WPM) image with 2 m spatial resolution. A U-Net convolutional neural network was developed using Python and the Keras and TensorFlow libraries. Ground truth data for training and accuracy assessment were generated through supervised classification of the same image in ArcGIS Pro, employing the Support Vector Machine (SVM) algorithm, followed by post-classification procedures, including majority filtering and manual pixel editing. U-Net results were compared with SVM results (pre-refinement) to evaluate potential accuracy improvements associated with the greater computational and human effort required by Deep Learning techniques. Both approaches were assessed using Overall Accuracy, Precision, Recall, F1 Score, and Kappa metrics. The U-Net model demonstrated superior performance across all metrics, with a notable increase in Precision from 0.48 (SVM) to 0.78 (U-Net). These findings highlight the potential of Deep Learning methods for high-resolution urban LULC mapping, providing valuable tools for urban planning and territorial management in Brazilian municipalities.