Pseudo-labeling-based change detection in SAR image time series using machine learning

This study aims to evaluate an automated approach for mapping areas affected by land use and land cover change using Sentinel-1 SAR imagery time series and machine learning models trained through a novel pseudo-labeling framework. The proposed framework automatically identifies reliable pseudo-labeled samples for training classifiers and represents an evolution of a previously proposed strategy developed for the same purpose. Different algorithms and attribute configurations were tested to automate the distinction between stable areas and regions subject to spatiotemporal changes. Two case studies were conducted in regions with distinct characteristics, located in Novo Progresso (Brazil) and Istanbul (Turkey), allowing a broad evaluation of the proposed method. The results indicate that the choice of classifier, the configuration of attributes, and regional characteristics influence the mapping performance. In Novo Progresso, the best results were obtained with SVM-Linear (kappa: 0.749; F1-Score: 0.842), followed by RF and RBFNet, while AdaBoost performed the worst. In Istanbul, RBFNet demonstrated higher accuracy (kappa: 0.759; F1-score: 0.852), confirming the effectiveness of the approach in detecting changes across different contexts.