Hyperspectral Object Detection via Spectral Prototype Mining and Spatial-Spectral Mamba Fusion

Abstract

Hyperspectral images record rich spectral responses across contiguous bands, providing fine-grained material cues for object detection beyond conventional visual appearance. However, the high redundancy among spectral bands and the complex spatial–spectral correlations pose significant challenges to effective feature representation and fusion. To address these issues, we propose a novel hyperspectral object detection (HOD) framework, called DSSM-HOD, that integrates a spectral prototype mining module (SPMM) with a spatial–spectral Mamba fusion module (SSFM). Specifically, SPMM is first introduced to discover representative spectral prototypes from redundant spectral bands, thereby preserving informative spectral signatures while suppressing redundant spectral responses. Subsequently, the SSFM is designed to jointly capture long-range spatial dependencies and spectral interactions through spatial and spectral Mamba modeling. Moreover, an adaptive fusion mechanism is employed to globally balance and integrate complementary spatial and spectral features from both domains. Experimental results on a large-scale HOD dataset demonstrate that the proposed DSSM-HOD consistently outperforms state-of-the-art methods, achieving superior detection accuracy and strong robustness across diverse scenarios.