LDMAN: A Lightweight Dynamic Dual-Path Convolution Assisted MLLA Network for Hyperspectral Image Classification
Abstract
Hyperspectral image (HSI) classification requires joint spectral–spatial feature learning, whereas modeling complex local–global dependencies often increases computational cost. To address this limitation, a lightweight dynamic dual-path convolution assisted Mamba-like linear attention network (LDMAN) is proposed. The network comprises two components: the gated Mamba-like linear attention (GMLLA) module and the dynamic dual-path convolution module (DDPCM). GMLLA models long-range dependencies with linear computational complexity using a global input gate (GIG) and improves nonlinear spectral feature representation under limited samples through parameter-efficient KANLinear projections. DDPCM adaptively routes spectral channels to a semantic path for high-level feature extraction or a spatial path for structural detail preservation under the guidance of an attention-based routing mechanism. Experimental results on two HSI datasets demonstrate that LDMAN achieves higher classification accuracy while requiring fewer parameters and lower floating-point operations (FLOPs) compared to state-of-the-art methods. The code is available at https://github.com/sg10216/LDMAN for reproducibility.