PyS2CF-Mamba: A Pyramid Spatial–Spectral Competitive Fusion Mamba Network for Hyperspectral Image Classification

Abstract

Hyperspectral image classification (HSIC) is a fundamental task in remote sensing scene understanding. However, existing methods still face the following three challenges: a balance between local edge details and long-range spatial dependencies, weak discriminative spectral dependency modeling, and a lack of adaptability in fusing heterogeneous spatial–spectral features. To address these issues, we propose a pyramid spatial–spectral competitive fusion Mamba network (PyS2CF-Mamba), which constructs a spatial–spectral dual-branch backbone to decouple feature learning and adaptively integrates complementary features through competitive fusion. In particular, the local-prior pyramid spatial Mamba (LPPS-Mamba) branch integrates a lightweight spatial prior (LSP) module, a pyramid refined channel attention (PRCA) module, and Mamba-based global scanning to jointly capture local edge details and long-range spatial dependencies. In parallel, the differential grouped spectral Mamba (DGS-Mamba) branch couples first-order latent feature-channel difference enhancement with grouped sequence scanning, enhancing channelwise variations in the compressed latent space. On this basis, a channelwise competitive fusion module dynamically integrates complementary features via adaptive competitive weighting, enhancing the adaptability of heterogeneous spatial–spectral feature fusion. Experiments on the WHU-Hi-LongKou (LK), QUH-Qingyun (QY), and QUH-Tangdaowan (TDW) datasets demonstrate that PyS2CF-Mamba achieves superior classification performance compared with convolutional neural network (CNN)-, Transformer-, and Mamba-based baselines. The code is available at https://github.com/JiaxinLiCAS