关键词: 树种分类, 小样本, 注意力机制, 自监督对比学习, 冠层高度信息

Abstract:

Objective: To address the challenges of small sample sizes, spectral confusion, and weak cross-sensor generalization in tree species classification, a prototype network based on self-supervised contrastive learning was constructed. By integrating satellite-borne multispectral data with airborne LiDAR elevation information, this study aims to explore its classification performance and robustness under small-sample conditions and across different sensors. Method: Based on field survey data and high-definition historical images from Google Earth, a sample dataset containing 3 major categories and 10 minor categories was constructed. Sentinel-2 and GF-6 satellite multispectral images were utilized respectively, and the airborne LiDAR was introduced to generate canopy height model (CHM) as an extended band, the height information was used to enhance the discrimination ability between vegetation and non-vegetation. An improved prototype network integrating the CBAM attention mechanism and SimCLR self-supervised contrastive learning was proposed to enhance feature representation and small sample learning ability. The control variable method was used to evaluate the impact of the band number and spatial resolution on tree species recognition, and verify the generalization performance of the classification model on different datasets. Result: The tree species classification accuracy on Sentinel-2 and GF-6 data both exceeded 85%. When the number of samples per category was controlled at 5–10, a relatively ideal classification effect could be achieved, indicating the robustness of the model under small sample conditions. The introduction of attention mechanism and self-supervised strategy improved the classification accuracy by more than 2.00% compared with the baseline prototypical network, and significantly outperformed Random Forest (RF) and two-dimensional convolutional neural networks (2D-CNN). Cross-sensor experiments revealed that Sentinel-2 offered superior spectral discrimination, while GF-6 with its 2 m resolution performed better in boundary delineation. The overlapping ratio of the two classification results generally exceeded 60%, confirming strong cross-sensor generalization ability. After combining LiDAR canopy height information, the overall accuracy (OA) of Sentinel-2 and GF-6 data increased by 3.85% and 2.73%, respectively, effectively alleviating misclassifications among spectrally similar species. Conclusion: The integration of satellite multispectral imagery and LiDAR data effectively overcomes the limitations of traditional two-dimensional spectral information classification. The improved prototype network significantly enhances small sample learning efficiency and cross-sensor generalization ability through the collaborative optimization of the attention mechanism and contrastive learning, providing a feasible technical path for large-scale forest resource dynamic monitoring.

Key words: tree species classification, small sample, attention mechanism, self-supervised contrastive learning, canopy height model