高光谱成像
遥感
接头(建筑物)
激光雷达
计算机科学
地质学
人工智能
环境科学
雷达跟踪器
上下文图像分类
作者
Jiangyun Li,Jiaheng Han,Hao Wang,Xiaochen Zhang,Peixian Zhuang
标识
DOI:10.1109/tgrs.2026.3703805
摘要
Although recent advances have improved joint hyperspectral images (HSIs) and light detection and ranging (LiDAR) data classification, most existing approaches adopt patch-based processing strategy that leads to substantial computational redundancy for high-resolution data. Moreover, existing architectures usually process spectral and spatial features independently, making them difficult to capture their relationships and thus limits the fusion performance. To address the above challenges, we propose WIMNet, a Whole-Image-Based Mamba Network that performs image-level fusion for joint hyperspectral and LiDAR classification. Unlike existing patch-based approaches, the proposed WIMNet consists of two dedicated branches: Joint Spectral–Spatial Mamba Branch (JSMB), incorporating a Spa–Spe Mamba Block (SSMB) with a Spectral Modulation Mechanism (SMM) for adaptive spectral–spatial joint learning, and the Multiscale LiDAR Mamba Branch (MLMB) for LiDAR feature extraction. Then their features are integrated via a Multimodal Feature Fusion Module (MFFM) to produce the final classification map. Final extensive experiments on Houston2013, MUUFL, and Trento datasets demonstrate that WIMNet achieves state-of-the-art accuracy while reducing GFLOPs and inference time by up to 99% compared with existing 11 leading methods, confirming its remarkable efficiency in joint hyperspectral and LiDAR classification. Code will be available at https://github.com/Dufu9/WIMNet.
科研通智能强力驱动
Strongly Powered by AbleSci AI