Discriminative Anchor Learning with Distribution Alignment for Multi-modal Remote Sensing Clustering

Anchor-based multi-modal clustering has emerged as a powerful paradigm for tackling complex data partitioning tasks, especially in large-scale scenarios such as remote sensing image (RSI) analysis. Despite the promising performance, existing anchor-based methods directly select the representative anchors for clustering, ignoring the distribution discrepancy between anchors and samples. This leads to the degradation of anchor quality and suboptimal clustering performance. To address the problem, we proposed a discriminative anchor learning with distribution alignment for multi-modal RSI clustering. Specifically, we incorporate both anchor learning and subspace graph construction into a unified optimization formulation. The two processes can guide each other to boost clustering performance. By sharing the anchor coefficient matrix, we reinforce the consistency between the inter-anchor relationships and the sample-anchor affinity relationships. Consequently, the learned anchors can capture the distribution structure of samples, achieving distribution alignment between anchors and samples. We further impose low-rank and probabilistic constraints on the consensus coefficient matrix, which effectively explores latent structural information and enhances anchors' discriminative ability. Extensive experimental results on public multi-modal RSI benchmarks demonstrate the superiority and effectiveness of the proposed method.