HOTMM: Direction-Aware Spatial-Spectral Mamba With Long-Range Memory Guidance for Hyperspectral Object Tracking

Hyperspectral object tracking (HOT) plays an increasingly important role in remote sensing surveillance, as rich spectral signatures in hyperspectral (HS) imagery enable more reliable target discrimination. However, existing HOT methods still face challenges in fully exploiting spatial-spectral-temporal interactions, including limited spatial adaptivity, insufficient spectral integration, and inadequate temporal context modeling. Although recent Mamba-based trackers offer effective sequence modeling for HOT, their predefined scanning paths with uniform aggregation tend to overlook the diverse directional structures in remote sensing scenarios. To address these challenges, this article proposes HOTMM, a novel Mamba-based framework that unifies direction-aware spatial-spectral modeling with long-range memory-guided temporal modulation to jointly capture directional spatial structures, interband spectral dependencies, and persistent target evolution. At the core of the framework is the mixture of directional Mamba experts (MDMEs) module, which leverages the mixture-of-experts (MoE) paradigm to transform such spatial-spectral cues into content-adaptive prompts. These prompts are then injected into RGB-pretrained tracking representations to enhance their discriminative capacity. Within MDME, we design a routed-and-shared expert scheme, where multiple directional spatial Mamba streams are sparsely activated and adaptively weighted based on the current spatial patterns, while bidirectional spectral scanning branches are shared to aggregate group-wise spectral features along the wavelength dimension. Furthermore, a temporal memory Mamba (TMM) mechanism is introduced to bridge deep intraframe representations with long-range interframe target evolution. By adopting a unidirectional causal scanning strategy, TMM generates memory-guided temporal modulation signals that continuously refine target descriptors under persistent appearance variations. Extensive experiments on eight public HS tracking datasets demonstrate that HOTMM achieves state-of-the-art (SOTA) tracking performance. The code will be available at https://github.com/creenciaxz/HOTMM