.png){kind=icon}
.png){kind=icon}
Return
AdaSFMNet: Adaptive Spatial–Frequency Mamba for HSI–LiDAR Joint Classification
.png){kind=icon}
PRE
AI
Z
H
N
L
A
DOI：10.1109/tgrs.2026.3716822.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The joint classification of hyperspectral image (HSI) and light detection and ranging (LiDAR) is an important task in remote sensing, as the complementary characteristics of the two modalities can significantly improve land cover recognition accuracy. However, existing methods mostly focus on spatial-domain feature modeling and lack sufficient exploration of frequency-domain information. Moreover, traditional wavelet transforms employ fixed filters, making it difficult to adaptively match the distinct frequency-domain characteristics of HSI and LiDAR. Meanwhile, these methods lack effective adaptation to frequency features, which limits fusion performance. To address these issues, this article proposes adaptive spatial-frequency mamba (AdaSFMNet), a joint classification network based on adaptive wavelet feature alignment and spatial–frequency Mamba (SF-Mamba). The proposed framework can adaptively extract spatial–frequency features while achieving global–local modeling across the spatial and frequency domains. For frequency-domain feature extraction, inspired by the discrete wavelet transform (DWT), learnable wavelet kernels are designed to adaptively separate the high-frequency and low-frequency components of HSI and LiDAR, and a cross-attention mechanism is employed to semantically align the high- and low-frequency components, thereby enabling fine-grained extraction and consistent representation tailored to the distinct frequency characteristics of each modality. To promote cross-domain fusion between the spatial and frequency domains, an SF-Mamba is introduced to fully exploit the spatial–frequency features of HSI and LiDAR. This design incorporates a spatial perception branch and a frequency filtering branch to complement the Mamba backbone with full spatial–frequency awareness, while also supporting both global and local modeling. Within AdaSFMNet, the adaptive wavelet decomposition and semantic alignment mechanism improve the extraction and alignment of multimodal features, while SF-Mamba removes redundant information through an adaptive spatial–frequency mapping branch and enhances the adaptation and fusion of cross-modal features in both the spatial and frequency dimensions. Their synergy produces more discriminative representations. Experimental results on three public datasets demonstrate the superiority of AdaSFMNet.
Keywords:
Adaptive wavelet transform
hyperspectral image (HSI)
joint classification
light detection and ranging (LiDAR)
spatial–frequency Mamba (SF-Mamba)
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W
Organization
Cited Papers