MVTopo: Multiview Topology-Aware Prototype Reconstruction for Generalized Category Discovery in Remote Sensing Images

Remote sensing image classification typically relies on closed-set assumptions. Generalized Category Discovery (GCD) overcomes this limitation by simultaneously classifying known categories and discovering novel ones in unlabeled data, going beyond out-of-distribution (OOD) detection and open-set recognition (OSR) that can only reject unknown samples without further categorization. However, GCD remains largely unexplored in remote sensing; directly transferring natural-image GCD methods faces two domain-specific difficulties: single-view feature representations lack the discriminability to handle severe intra-class variation and inter-class spectral similarity, while indiscriminate clustering mechanisms are vulnerable to boundary noise in complex backgrounds, leading to degraded known-class recognition and false-positive assignment of novel categories. To address these challenges, we propose MVTopo, a Multi-View Topology-aware framework for remote sensing GCD. MVTopo decouples single-modal features into semantic, spectral, and spatial views through lightweight attention modules, namely Global Low-rank Spectral Learning (GLSL) and Orthogonal Semantic Agent (OSA), to construct discriminative multi-view prototypes. A Topology-Aware Contrastive Mean Shift (TA-CMS) mechanism further refines clustering by dynamically weighting neighbors based on manifold topological confidence, suppressing boundary noise interference and effectively driving boundary samples toward their respective cluster centers. Experiments on three datasets, including AID, RSSDIVCS, and a newly constructed 90-class dataset DivRS-90, show that MVTopo achieves the best All and Novel accuracy on all benchmarks, reaching 90.75%/89.06% on DivRS-90, 91.72%/89.14% on RSSDIVCS, and 86.84%/84.16% on AID, surpassing the best competing methods by up to 2.49% on All and 1.83% on Novel.