Spatio–Temporal Clustering and Active Learning for Change Classification in Satellite Image Time Series

Abstract

Active learning (AL) has emerged as a versatile approach to reduce the training data required for remote sensing image classification, but its use for the analysis of satellite image time series (SITS) has not been explored yet. This study targets to explore a new object-based framework for change detection in SITS, combining the state-of-the-art spatio-temporal clustering and the use of different machine learning algorithms. Indeed, this study aims at testing whether standard machine learning algorithms can detect changes in long time series and whether AL can improve the results compared to traditional supervised learning. The tested AL algorithms comprise random forest-based heuristics that use a combination of uncertainty and diversity criteria, and a classical SVM breaking ties heuristic. The different implementations are evaluated with two datasets that depict changes around the Arcachon basin (West of France) and around the city of Colmar (East of France) spanning over more than 20 years and comprising four change classes (urban sprawl, forest gain, forest loss, and other). The tests demonstrate that steeper learning curves can be obtained with AL when compared to supervised learning. However, the performance of different AL algorithms depends on the dataset.