Multisource Adaptive Feature Enhancement and Fusion Network for Coastal Wetland Classification With Hyperspectral and LiDAR Data

Wetland classification via remote sensing is crucial for ecological protection. In coastal wetlands, severe spectral similarity between different vegetation and strong spectral variability of the same vegetation due to growth environment differences present challenges. Integrating hyperspectral image (HSI) with rich spectral details and light detection and ranging (LiDAR) data offering precise height information shows significant potential, yet effectively extracting complementary features and achieving efficient fusion remains a key issue in such classification. This paper proposes a multisource adaptive feature enhancement and fusion network (MsAF-EFN), comprising three core components: feature extraction, enhancement, and fusion. To explore key features of the two data sources, an asymmetric dimensional scaling module is designed for efficient extraction of spectral and height information. Subsequently, spectral adaptive attention and direct height-aware channel attention modules are developed to optimize features and capture discriminant semantic information reflecting subtle differences of ground objects in higher-order features. Finally, a dynamic spectralspatial-height integration module adaptively calculates feature importance while preserving low-frequency residual information, enabling efficient integration of multisource features. To verify MsAF-EFN and supplement the scarce public datasets in the field of wetland classification research, we independently constructed a wetland dataset containing HSI and LiDAR data from two years in the Yellow River Delta. Experiments on two coastal wetland datasets confirm that MsAF-EFN outperforms state-of-the-art methods, particularly in scenarios with high spectral similarity and complex spatial structures, validating its effectiveness in multisource data wetland classification. The dataset and code will be publicly available upon paper acceptance at https://github.com/UPCGIT/MsAF-EFN.