• The study proposes Mangrove–Net, a point-wise DL model for mangrove species mapping. • Integrates multi-stage attention & class token for local and global feature balance. • Our method outperforms 6 benchmark models (OA:75.26%, mAcc:74.79%, min + 6.6%/+7.7%). • Multi-scale visualization shows our species maps with better ecological consistency. • Provides a reliable method for fine-grained mangrove species classification. Mapping mangrove species is crucial for mangrove biodiversity protection and ecological restoration. Existing mangrove classification methods mostly rely on the height and intensity metrics of rasterized lidar point clouds, which lead to the partial loss of fine three-dimensional (3D) structural information, and are difficult to distinguish among small inter-class differences, complex structures (dense canopy and aerial roots), and uneven mangrove stands. To directly capture the species community structure of mangroves from a 3D perspective, this study develops a point-wise deep learning species classification method specifically for mangroves—called Mangrove Multi-stage Attention Fusion and Class Token Network (Mangrove-Net), using UAV lidar point clouds and RGB imagery fusion data. Mangrove-Net uses a multi-stage attention fusion and class token mechanism to balance local geometric representation with global semantic perception, effectively capturing representative features among mangrove species. To address sample-imbalanced classification tasks, Mangrove-Net introduces a weighted cross-entropy loss function. We tested the proposed method on a typical mangrove forest dataset (from the Hainan Qinglangang Provincial Nature Reserve, China) and on the open dataset ModelNet40. The results indicated that Mangrove-Net showed superior performance compared to six state-of-the-art point-based deep learning methods (such as PointNet++, Point Transformer, PointMLP, and GPSFormer) in mapping fine-grained mangrove species, with a minimum improvement of 6.60% and 7.71% in overall accuracy (OA) and mean category accuracy (mAcc), respectively. Multi-scale visualization (sample, local, and study area) showed that the mangrove species maps generated by our proposed method had higher spatial continuity and ecological consistency. Evaluated on the ModelNet40 dataset, Mangrove-Net also surpassed those baseline methods, further demonstrating the effectiveness and robustness of the method. This study provides a reliable point-wise deep learning method specifically for mangrove species classification to support various tasks in fine-grained precision mangrove forestry.
Wang et al. (Fri,) studied this question.