关键词: 遥感影像, 树种分类, 单域泛化, 语义分割, 轻量化模型

Abstract: Objective This study aims to address domain shifts caused by regional variations in ecological factors, such as climate and soil, for which a single-domain generalization method was proposed based on global-local feature fusion, so as to enhance the generalization performance of unlabeled tree species recognition in complex forest scenes and provide new method support for cross-regional tree classification research. Method The southern part of Baden-Württemberg, Germany, and the western part of Qimen County, Huangshan City, China, were selected as the source domains, while the central part of Thuringia, Germany, and the eastern part of Qimen County, Huangshan City, China, were selected as the target domains. A global-local feature fusion network (hierarchical unified feature network, HUFNet) was constructed for tree species classification. This network consists of a CNN-based encoder layer, a Transformer-based decoder layer, a Global–Local Attention Feature Extraction (GLAFE) mechanism, a Feature Refinement Head (FRH), and an Edge Refinement and Validation (ERV) module. The model was trained on the source domain datasets and then tested on the target domain to validate its generalization ability, achieving cross-domain tree species classification in complex scenarios.Result By comparing multiple source and target domain datasets, the HUFNet model achieved an overall accuracy (OA) of 75.1% and a mean intersection over union (mIoU) of 58.3% for the classification of coniferous and broadleaf tree species on the target domain HainichUAV dataset. Compared to the classification architecture based on self-attention mechanisms, the model improved OA and mIoU by 13.7% and 11.7%, respectively. On the HuangshanEast target domain dataset, the HUFNet model achieved OA of 71.7% and mIoU of 56.8%. Compared to the hybrid architecture using ViT-R50 as the encoder, the OA was improved by 1.2%. Conclusion The HUFNet model proposed in this study achieves significant improvements in cross-regional tree species classification, it not only maintains the high-precision recognition ability, but also shows the powerful cross-domain generalization ability in the target domain, and greatly reduces the time and space complexity of the model. It shows strong application potential in resource-constrained environments. The single-domain generalization approach based on global-local feature fusion provides a novel perspective for cross-domain tree species classification in advancing research.

Key words: remote sensing imagery, tree species classification, single-domain generalization, semantic segmentation, lightweight model