Disentangling Complex Forest Canopies Using MS 3 Mamba With Hierarchical HSI-LiDAR Fusion for Robust Tree Species Classification
Chao Li, Shijie Wang, Wenjun Zhang, Ying Quan, Yuan Zhou, Guofan Shao, Ye Yuan, Weipeng Jing
Northeast Forestry University Purdue University West Lafayette
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Accurate tree species classification is critical for sustainable forest management, but is hindered by the high intraspecies variability and low interspecies separability, inherent in hyperspectral imaging (HSI) data. While the fusion of HSI with light detection and ranging (LiDAR) data offers a promising solution, existing fusion methods exhibit a critical limitation, as they are often designed for general scenes with high interclass variance, and thus are unable to resolve the subtle spectral ambiguities of complex, occluded forest canopies, leading to significant pixel-level misclassifications. To enhance the separability of tree species with high spectral similarity and to mitigate the resulting misclassifications, we propose multisource spectral–spatial Mamba (MS3Mamba), an end-to-end hierarchical fusion architecture specifically engineered for tree species classification. MS3Mamba strengthens feature learning by fusing shallow and deep features extracted by the spectral-structural feature encoding (SSFE) and global canopy context modeling (GCCM) modules, respectively. It synergizes local features from a multiscale convolutional neural network (CNN) with global features from spatial and spectral Mamba blocks for complementary integration, utilizing two independent modal data streams to avoid mutual interference. Furthermore, the tree-aware detail refinement (TADR) module optimizes the spatial–spectral fused features at tree canopy boundaries to enhance classification accuracy. To rigorously validate model robustness, we establish the MNFPTree-HL dataset, a new benchmark from a structurally complex natural secondary forest, comprising co-registered, submeter HSI (0.3 m) and LiDAR (0.1 m) data with an extensive ground-truth inventory (709 trees, 11 species) that addresses the critical research gap for evaluating the performance under these challenging conditions. Comprehensive experiments demonstrate that MS3Mamba achieves state-of-the-art (SOTA) performance by effectively disentangling spectrally similar species and accurately discriminating complex canopy boundaries, particularly on our new MNFPTree-HL benchmark.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
生物医学Remote Sensing and LiDAR Applications
Remote-Sensing Image Classification · Advanced Image Fusion Techniques
参考文献 43
此处列出前 3 条