Abstract
Most existing studies on change detection (CD) focus on binary change detection, which identifies changed regions but does not distinguish different types of land-cover change. Unsupervised multiclass change detection provides this additional information without requiring manually annotated training samples. However, the number of change categories is often unknown, and spectral variability and overlapping change responses can lead to incorrect cluster assignments or redundant categories. Using these clustering results as pseudo-labels can propagate errors into subsequent network training. Moreover, pixel-wise spectral differences alone may be insufficient to separate change categories, making spatial context necessary for their discrimination. To address these issues, we propose an adaptive pseudo-label-guided cross Mamba framework (APC-Mamba) for unsupervised multiclass remote sensing CD. An adaptive pseudo-label generation and refinement module is developed to estimate potential change categories and provide reliable supervision by generating high-quality pseudo-labels. Meanwhile, a Wavelet Mamba Residual Block (WMRB) is introduced to model low-frequency spatial context, which is then used to guide the refinement of high-frequency detail features; this helps preserve the overall structure and local boundaries in regions of change, while simultaneously suppressing interference from high-frequency noise. Furthermore, a Spatial-Channel Cross Mamba module (SCCM) is designed to facilitate feature interaction and information fusion across spatial and channel dimensions, producing more discriminative bitemporal representations. The experimental results demonstrate the effectiveness of APC-Mamba across hyperspectral and multispectral datasets, particularly in terms of overall accuracy and Kappa coefficient.