Revisiting the Effectiveness of Self-Supervised Learning for Change Detection in Remote Sensing Imagery

Abstract

Remote sensing image change detection is a core task in Earth observation and geographic information science. In recent years, deep learning-based change detection methods have become the mainstream approach. However, these fully supervised methods impose stringent requirements on both the quantity and quality of labeled data. Consequently, self-supervised and unsupervised methods, which reduce or eliminate dependence on manual annotations, have attracted considerable research attention. Nevertheless, this study finds that, within a typical change detection paradigm combining self-supervised pretraining with unsupervised postprocessing, the self-supervised learning process fails to provide effective optimization. Its training behavior is comparable to random initialization and may even lead to performance degradation, exhibiting characteristics of a pseudorandom process. Experiments on the publicly available OSCD and AirChange datasets support this finding. Further analysis shows that self-supervised contrastive change detection methods can still achieve a certain level of performance because their feature extraction network, ResUNet, preserves low-level visual information. We argue that these low-level visual features, rather than the self-supervised learning process itself, are the key factor underlying the effectiveness of this class of methods. To further validate this argument, we replace the feature extraction networks used in the investigated paradigm with a single convolutional layer. This simplified approach achieves the best performance within the examined paradigm, obtaining F1-scores of 0.529 and 0.263 on the OSCD and AirChange datasets, respectively, outperforming the best-performing baseline on each dataset by 0.2 and 3.5 percentage points.