A Multi‐View Tree Species Classification Method Based on the Attention Mechanism

Tree species information is vital for forest resource surveys and serves as foundational data for forest resource management. Terrestrial Laser Scanning (TLS) can quickly capture the three‐dimensional structural information of trees with high resolution. Tree species can be automatically and accurately identified from complex individual tree point clouds using deep learning methods, which is a key research direction for TLS in forestry applications. To enable deep learning models to fully extract and learn the feature representations of trees, this study addresses the limitations of existing methods in tree species classification and proposes a multi‐view tree species classification method based on the attention mechanism. Given an individual tree point cloud, this study uses six camera projections to generate six three‐channel images based on depth information instead of six single‐channel depth images as input to the network, fully simulating the observed change in point cloud distance information from a certain direction. In addition, the six sets of output features from the network are input to the attention module to learn global features instead of being simply concatenated to represent global features. The attention module focuses on the dependencies between the features of six view images and can extract more efficient classification features. The effectiveness of the proposed method was evaluated using the Wytham Woods dataset and Mediterranean forests, with performance assessed in terms of accuracy and the kappa coefficient. Experimental results demonstrate that the proposed method achieves state‐of‐the‐art performance, showcasing its potential for advancing tree species classification in forestry applications.