Switcher-HNet: A switchable hierarchical network for tree species classiffcation from forest stand to individual tree tasks
Date:2025-10-31 clicks:
DOI number:10.1016/j.isprsjprs.2025.10.040
Affiliation of Author(s):东北林业大学
Journal:ISPRS Journal of Photogrammetry and Remote Sensing（Q1,TOP）
Key Words:Tree species classiffcation; Individual tree; Forest stand; Deep learning; Transfer learning
Abstract:Tree species information is essential for effective forest management, ecosystem health, and species biodiversity. Currently, most classiffcation models treat tree species as independent categories and overlook the hierarchical relationships among these species, hindering the accuracy and effectiveness of tree species classiffcation. To address this issue, this study proposes Switcher-HNet, a switchable hierarchical convolutional network integrating a backbone network, a hierarchical module, and a switcher mechanism to leverage tree species hierarchies and enhance their identiffcation. This study introduces two hierarchical datasets structured at three levels: land cover (level 0), foliage type (level 1), and tree species (level 2). The TreeSatAI dataset, collected in Germany, is designed for forest stand classiffcation with dominant species such as pine, beech, spruce, and oak. The Hierarchical Individual Tree (HIT) dataset, acquired in northeastern China, targets individual tree classiffcation and is primarily composed of Pinus koraiensis and various broadleaf species. These two datasets provide a structured label hierarchy and a multi-scale perspective across biogeographically distinct regions. SENet, the topperforming backbone among the evaluated networks, was selected as the shared backbone for both SwitcherHNet and a hierarchical baseline on these datasets. Results show that Switcher-HNet outperformed the baseline on the TreeSatAI dataset, achieving a 0.89% improvement in overall accuracy (OA) at level 1 and a signiffcant 2.95% increase at level 2. To evaluate transferability, Switcher-HNet, pretrained on the TreeSatAI dataset, was ffne-tuned on the HIT dataset for individual tree classiffcation. Switcher-HNet achieved high accuracy with OAs of 94.00% at level 1 and 90.59% at level 2 on the full HIT dataset, reducing training time to less than one-third of that required for training from scratch. Therefore, the ffne-tuning techniques enable a smooth transition from forest stand to individual tree classiffcation. This study emphasizes the potential of hierarchical classiffcation methods to enhance tree species identiffcation and provides a technical support for sustainable forest management and biodiversity conservation.
Co-author:Xiang Li
First Author:Shengheng Liu
Indexed by:Journal paper
Correspondence Author:Zhen Zhen*,Yinghui Zhao*
Volume:231
Page Number:329-344
Translation or Not:no
Date of Publication:2025-01-01
Included Journals:SCI