Research on Forest Tree Species Classification Via Unmanned Aerial Vehicle (Uav) Optical Imagery and Unsupervised Methods

Abstract

Tree species information is a critical foundation for forest ecological research, with applications spanning forest ecosystems, biodiversity assessment, and conservation efforts. Unmanned aerial vehicles (UAVs) are critical tools for forest data collection, yet cross-domain shifts pose significant challenges to dataset construction and ecological data comprehensiveness. These shifts arise from differences in data distributions across domains, driven by seasonal variations, illumination conditions, sensor types, and regional differences. Tree species exhibit color and morphological changes due to seasonal or climatic factors, while variations in lighting, sensors, or backgrounds can disrupt image consistency. In addition, variations in tree species distribution across geographically distinct areas pose challenges for model generalization when trained on a single dataset. Existing solutions to cross-domain shifts are often costly or complex. Supervised learning requires extensive annotated data to address distribution inconsistencies, which is time-intensive and expensive. Unsupervised learning eliminates the need for annotations but suffers from cluster instability and limited robustness. Transfer learning and synthetic data offer partial solutions but fail to fully capture real-world feature diversity. To address these challenges, this study proposes an innovative unsupervised learning approach integrating pretrained models, UMAP dimensionality reduction, and clustering voting algorithms. Experimental results demonstrate the method's effectiveness: it achieves 85.1% classification accuracy on a public dataset of 29,652 images and 91.67% accuracy in field experiments using UAV-collected data, outperforming existing supervised learning methods. This approach reduces annotation costs, streamlines dataset construction, and provides a robust solution for large-scale forest monitoring and ecological research.

Keywords: Unmanned Aerial Vehicle (UAV) Remote Sensing, Optical Imaging, Tree Species Classification, Unsupervised Methods