GMambaHSI: Group-based visual state space model for hyperspectral image classification

Hyperspectral image (HSI) classification is essential in the remote sensing (RS) domain. Transformers have gained prominence in this domain in recent years owing to their capacity for global information modeling. However, the quadratic complexity constrains their efficacy when computer resources are restricted. A structured state space model (SSM) called Mamba has emerged. Like Transformers, it excels at modeling long-range dependencies in latent data. However, unlike Transformers, its complexity is linear. As a result, recent research has increasingly explored its efficacy in HSI classification. Nonetheless, the majority of them apply Mamba exclusively to HSIs without adequately considering the intrinsic properties of HSIs. This article introduces a novel, parameter-efficient modulated group Mamba layer designed to fully leverage Mamba's capabilities in HSI classification. It segments the input channels into four groups and independently applies the proposed SSM-based efficient visual single selective scanning (VSSS) block to each group, with each VSSS block scanning in one of four spatial orientations. The Modulated Group Mamba layer (MGML) encapsulates the four VSSS. In addition to MGML, an efficient Multi-kernel depthwise convolution (MK-DeConv) is also used. The model's backbone is an encoder block that uses MK-DeConv and MGML. By reducing unnecessary parameters and calculations, channel grouping and multi-kernel depthwise convolutions make things more efficient while still capturing both fine-grained and large-scale features. This combination makes it possible to create models that are both light and expressive while still being very accurate and not using too much memory or processing power. Extensive analyses of five benchmark hyperspectral imaging datasets (Pavia University, Houston, HanChuan, HongHu, and LongKou) reveal that GMambaHSI achieves enhanced classification accuracy while employing significantly fewer parameters compared to current CNN-, Transformer-, and Mamba-based models. In the Pavia University dataset, GMambaHSI, for example, raises the overall accuracy (OA) from 95.74% to 97.22% and the average accuracy (AA) from 95.8% to 97.87%.