Haimiao Ge

Accurate hyperspectral image (HSI) classification remains challenging, as discriminative performance critically depends on the effective coordination of high-dimensional spectral information and complex spatial structures. Recent deep learning methods, particularly attention mechanisms, have improved HSI classification. However, existing models still face spectral over-smoothing, spatial ambiguity, and poor spectral-spatial coordination. To address these limitations, a multi-scale spectral-spatial coordinated awareness fusion network (MS²CAFNet) is proposed for HSI classification. The proposed framework adopts a dual-branch architecture to explicitly decouple spectral and spatial feature learning, thereby alleviating cross-domain interference. In the spectral branch, a difference-gated multi-scale spectral attention (DGMSA) module is designed to jointly model spectral similarity and difference cues under multiple receptive fields, enabling an adaptive balance between redundancy suppression and fine-grained discrimination. In the spatial branch, a multi-head coordinate-aware spatial attention (MHCASA) module is introduced, which integrates explicit positional encoding with long-range contextual modeling to enhance robustness in spatially heterogeneous scenes. Furthermore, a bidirectional cross-guided fusion (BCGF) strategy is developed to align and coordinate spectral and spatial representations through correlation-based interaction, ensuring semantic correspondence across branches. Experimental results on four publicly available HSI datasets demonstrate that the network outperforms state-of-the-art methods.

With advances in deep-space exploration, hyperspectral imaging of the Martian surface has become feasible, making surface material classification a key tool for Mars observation. However, unlike terrestrial scenes, Martian mineral classes exhibit subtle spectral differences and highly similar spatial structures, resulting in strong spectral-spatial similarity among different materials and significantly limiting the effectiveness of existing methods. To address these issues, we propose a hybrid multi-hop differential Transformer with a pyramid gating mechanism (HMDTPG) for Mars hyperspectral image classification. The key novelty lies in the multi-hop differential Transformer, which explicitly models spectral differences via a multi-hop differential attention mechanism, enhancing sensitivity to subtle mineral variations while preserving global spectral dependencies. In parallel, a spatial pyramid gated enhancement module is designed to capture discriminative features across multiple scales, suppress redundant information, and retain fine-grained discriminative spatial features through gating. Furthermore, a spectral-spatial adaptive fusion module is designed to enable deep interaction between spectral and spatial features and perform adaptive fusion. Experimental results show that HMDTPG outperforms other state-of-the-art methods on three public MHSI datasets, achieving overall accuracies of 98.32%, 97.34%, and 97.42% on Holden Crater, Nili Fossae, and Utopia Planitia datasets, respectively.

The fusion of hyperspectral images data (HSI) and Light Detection and Ranging data (LiDAR) has long been a research hotspot in remote sensing. Nevertheless, most existing fusion methods directly take HSI and LiDAR data as model inputs, without adequately addressing the spectral redundancy caused by the high-dimensional correlations inherent in HSI, or the cross-modal spatial redundancy induced by structural overlap in complex HSI and LiDAR datasets. In addition, although spectral, spatial, and other attribute features have been extensively explored, the latent and complex relationships among features remain insufficiently modeled. To overcome these, this paper proposes a coarse-to-fine collaborative representation network (CFCRNet). In the coarse-grained feature learning stage, Spatial Reconstruction Module models the discriminability score of spatial positions within a single channel and derives a spatial attention weight formula based on this score, which can be reused across all channels for spatial information reconstruction. Spectral Redundancy Removal Module integrates spectral grouping and spectral normalization to reduce global spectral band correlations, and focuses on suppressing redundant components in local spectral variations, thereby facilitating efficient subsequent feature fusion. In the fine-grained feature learning stage, Multi-Head Factorized Attention Fusion Module focuses on selecting and fusing common and distinctive features, and explicitly models feature dependencies to enhance the discriminability of fused representations. Based on an integrated “redundancy removal-feature learning” mechanism, the proposed model can capture more effective complementary features under complex land-cover backgrounds. Experimental evaluations on three publicly available datasets show that CFCRNet outperforms state-of-the-art methods in land-cover classification tasks.

Pixel-level land cover classification (LCC) is a fundamental and challenging task in Earth observation. Recently, the integration of hyperspectral imagery (HSI) and light detection and ranging (LiDAR) data has significantly enhanced classification accuracy by leveraging their complementary spatial characteristics. Most existing LCC methods predominantly operate in the spatial domain. Although spatial-domain modeling preserves explicit spatial patterns, it often neglects latent correlations embedded in the frequency domain. Moreover, treating spatial- and frequency-domain representations independently is insufficient for modeling the complex dependencies in heterogeneous HSI–LiDAR data. In this article, we propose the frequency-driven state space model with cross-scale spatial aggregation (FSSM-CSA), which comprises three major components: the frequency-guided hierarchical feature collaboration (FG-HFC) module, the cross-scale spatial aggregation (CSSA) module, and the frequency–spatial adaptive fusion (FSAF) module. Specifically, the FG-HFC module embeds learnable frequency encodings into the Mamba block within a hierarchical structure, where residual connections and adaptive interlayer aggregation promote effective feature collaboration. The CSSA module captures cross-scale spatial dependencies through optimized downsampling and a novel tree-structured fusion strategy, enabling complementary integration of heterogeneous HSI and LiDAR features. Finally, the FSAF module adaptively balances the contributions of spatial and frequency features via a dynamic weighting mechanism, yielding more discriminative and complementary representations. Extensive experiments on three public benchmark datasets verify that FSSM-CSA consistently outperforms existing methods, validating its effectiveness in jointly modeling spatial and frequency information and in extracting discriminative features from heterogeneous data.

With the rapid development of Earth observation technologies, the integration of multi-source remote sensing data from diverse sensors offers greater potential for remote sensing image classification. Hyperspectral images (HSI) provide rich spectral and spatial information, while light detection and ranging (LiDAR) offers precise three-dimensional structural data. The effective fusion of these two modalities can significantly improve the accuracy of land cover classification. However, existing methods often preserve considerable redundant information during feature extraction and fusion, and the interaction among this redundant information significantly impairs the discriminative capability of the network. In addition, due to the irregular spatial distributions of multi-source remote sensing images, rule-based meshing methods often fail to model complex geometric structures and spatial dependencies, which reduces the ability of the network to perceive subtle spatial features across domains. To address the aforementioned challenges, a multilevel differential aggregation with gated discriminative network (MDAGNet) based on dynamic feature selection is proposed for joint land cover classification of HSI and LiDAR data. First, a multilevel differential aggregation module (ML-DAM) is proposed to extract spectral features from HSI and elevation features from LiDAR data separately, thereby generating more discriminative multilevel feature representations. Second, we meticulously design a gating mechanism-based adaptive discrimination module (ADM). This module adaptively distills useless redundant information through multi-scale gated discriminative operations, enabling deep fusion of critical detail features. Furthermore, we introduce a morphological strip convolution module (Morpho-SConv) to characterize the complex geometric spatial distribution in HSI, enhancing the ability of the network to perceive fine-grained features. Finally, we achieve adaptive fusion of multi-modal data through weight coupling and dynamic fusion mechanisms. Extensive experiments on three public datasets demonstrate that our method outperforms state-of-the-art approaches in both accuracy and generalization capability.

In recent years, with the continuous advancement of Earth observation technologies, the joint utilization of hyperspectral imagery (HSI) and light detection and ranging (LiDAR) for classifying complex land cover types has garnered widespread attention in the field of multi-source remote sensing. However, the diversity of land cover increases the complexity of the spatial and spectral structures in remote sensing data, creating challenges for extracting discriminative features effectively. Moreover, the heterogeneity of multi-source remote sensing data poses challenges for existing methods to effectively integrate complementary information from various data sources. To address these challenges, a multi-scale split-recombination cooperative fusion network (MSRCFNet) is proposed for joint land cover classification using HSI and LiDAR data. It consists of three main components: multiple parallel multi-scale hierarchical inverted-pyramid (MHIP) modules, a cross-modal cooperative fusion module (CCFM), and a classification module. The MHIP module comprises multiple convolutional split-recombination blocks (CSRBs) at various scales, designed to extract and fuse discriminative multi-scale features. CCFM leverages spatial-scale consistency and the self-attention mechanism to model global relationships between different modalities, and then applies the cross-attention mechanism to effectively integrate complementary information from heterogeneous data. The classification module transforms the fused features into the final classification results. Experimental results on three publicly available HSI-LiDAR datasets demonstrate the superiority of the proposed network over state-of-the-art methods.

With the rapid advancement of deep learning technologies, the use of computers for processing and analyzing multimodal remote sensing data has attracted significant attention. Joint land cover classification (LCC) using hyperspectral image (HSI) and light detection and ranging (LiDAR) data is one such key area of research. Recently, the emerging deep learning framework Mamba has shown superior performance over traditional architectures, including transformers and convolutional neural networks. However, its application to LCC faces challenges. First, existing methods typically use a unified feature extraction approach for heterogeneous data, which fails to capture the distinct characteristics of each data modality. Moreover, effectively integrating these heterogeneous features remains a significant difficulty. To address these issues, we propose the multimodal fusion Mamba network (M2FMNet), which consists of three key components: the spatial–spectral adaptive Mamba, the elevation-enhanced Mamba, and an enhanced fusion module. The spatial–spectral adaptive Mamba combines two distinct scanning methods to efficiently extract spatial and spectral features from HSI data. The elevation-enhanced Mamba, designed for LiDAR data, utilizes a specialized scanning method to extract elevation features. Finally, the enhanced fusion module integrates intra- and intermodal relationships of these features at the pixel level. Experimental results on three real-world datasets show that the M2FMNet outperforms state-of-the-art methods in both qualitative and quantitative metrics.