An Adaptive Pseudo-Labeled Sample Generation Approach to Unsupervised Multiclass Change Detection in Bitemporal Remote Sensing Images

Multi-class change detection (CD) in multitemporal multi-/hyperspectral remote sensing images is a significant yet challenging task, particularly when prior knowledge and ground reference data are not available. This paper proposes a novel adaptive pseudo-labeled sample generation (APSG) approach to address this challenge, which bridges the gap between data-driven unsupervised CD and supervised CD methodologies. In particular, the spectral difference image is first projected into a two-dimensional polar domain using an adaptive spectral change vector representation (ASCVR). Subsequently, a novel method for generating multiple pseudo-labeled candidate sample selection regions is developed to automatically analyze the statistical distribution and characteristics of change and no-change classes in bi-temporal images, thereby producing high-confidence pseudo-labeled samples. Finally, a sequential pseudo-labeled sample selection and generation strategy incorporated into the CD process is designed to enable robust multi-class CD. Multi-class CD results are further enhanced by incorporating spatial-spectral information. The effectiveness of the proposed method has been validated on three publicly available bi-temporal remote sensing datasets, outperformed the traditional CD approaches and the deep learning-based unsupervised multi-class CD methods in terms of the improvement of overall accuracy (1%–5%). The generated pseudo-labeled samples represent intrinsic change clusters and their feature distributions within the dataset, thus resulting in a data-driven approach. These pseudo-labeled samples serve as critical inputs for both machine learning- and deep learning-based CD methods utilizing multi/hyper-spectral data at medium-to- low spatial resolution. This approach enables reliable multi-class CD across diverse land-cover scenarios without requiring priori information.