WorldMamba: World-State Representation Learning for Hyperspectral Image Classification

Abstract

Existing hyperspectral image classification (HSIC) methods primarily learn discriminative spectral–spatial representations and directly map them to semantic labels, often overlooking the underlying material composition, scene organization, and physical characteristics of hyperspectral observations. To address this limitation, this work proposes WorldMamba (WMM), a novel framework for learning world-state representations for HSIC. WMM formulates HSIC as a structured world-state inference problem, where semantic predictions are derived from a latent world state rather than from a single discriminative feature embedding. Specifically, a 3-D spectral–spatial tokenizer first converts hyperspectral patches into compact tokens, which are subsequently processed by a vision Mamba world encoder to construct a unified world-state representation. The learned world state is explicitly decomposed into complementary material, scene, and physical states, capturing spectral composition, spatial organization, and intrinsic physical variations of hyperspectral scenes, respectively. The resulting world states are adaptively fused to produce the final classification. The code will be available at https://github.com/mahmad000/WorldMamba