SpecSpatMamba: an efficient hyperspectral image classification method integrating spectral-spatial dual-path and state space model

Current hyperspectral image classification methods face three critical limitations: (1) traditional CNNs suffer from the curse of dimensionality when processing high-dimensional spectral data, leading to overfitting and poor generalization; (2) existing approaches fail to effectively address spectra...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:The Egyptian journal of remote sensing and space sciences Ročník 28; číslo 4; s. 628 - 644
Hlavní autoři: Liao, Jianshang, Wang, Liguo
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier B.V 01.12.2025
Témata:
Deep learning
Hyperspectral image classification
Linear computational complexity
Spectral-spatial dual-path
State space model
Deep learning
State space model
Linear computational complexity
Spectral-spatial dual-path
Hyperspectral image classification
ISSN:1110-9823
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Current hyperspectral image classification methods face three critical limitations: (1) traditional CNNs suffer from the curse of dimensionality when processing high-dimensional spectral data, leading to overfitting and poor generalization; (2) existing approaches fail to effectively address spectral band redundancy, resulting in computational inefficiency and suboptimal feature representation; (3) conventional methods lack synergistic utilization of spatial-spectral information, treating spectral and spatial dimensions uniformly rather than exploiting their distinct characteristics. To address these gaps, this paper proposes SpecSpatMamba, a novel hyperspectral image classification method integrating spectral-spatial dual-path feature extraction with state space models. SpecSpatMamba introduces three core innovations: (1) Dual-path feature extraction with spectral-spatial separation, where 1 × 1 convolutions extract spectral features and 3 × 3 convolutions capture spatial features; (2) Hybrid architecture combining state space models with convolutional operations for balanced long-range dependency and local feature capture; (3) Computational efficiency breakthrough achieving O(L·d) linear complexity compared to Transformer’s O(L2·d) complexity. Experiments on four benchmark datasets—Indian Pines, Pavia University, Salinas Valley, and Houston2013—demonstrate competitive performance compared to state-of-the-art methods. SpecSpatMamba achieves overall accuracies of 95.11 %, 98.61 %, 96.97 %, and 91.48 %, respectively. Notably, SpecSpatMamba demonstrates superior cross-dataset consistency and robust performance across diverse geographic environments, with particularly strong improvements in complex urban scenarios (+0.39 % on Houston2013) and agricultural settings (+0.57 % on Salinas Valley), confirming the method’s effectiveness in addressing high-dimensional hyperspectral data challenges.
ISSN:1110-9823
DOI:10.1016/j.ejrs.2025.10.001