Tree Species Classification Using Majority Voting Approach on UAV Raw Images

Abstract. In recent years, the widespread availability of relatively low-cost cameras and unmanned aerial vehicles (UAVs) has made it easier to acquire high-resolution images of forests rich in texture information. In forest resource surveys, a common approach for tree species classification involves applying Structure from Motion (SfM) to the collected raw image set and using the resulting orthophotos for classification through deep learning techniques. This study proposes a novel tree species classification framework that aims to improve accuracy by utilizing UAV-acquired raw images, which are of high quality, can be captured in large volumes, and include diverse viewing angles. In the proposed method, tree species classification is first performed on each aerial image using a convolutional neural network (CNN) and a sliding window approach. Next, SfM processing is applied to the image set to generate a 3D point cloud and orthophotos. The classification results from the aerial images are then projected onto the point cloud, and finally, these projected results are mapped onto the orthophoto and aggregated to derive the final species classification by majority voting approach. The effectiveness of the proposed method is validated through experiments targeting four representative tree types in Japan: Cryptomeria japonica (Japanese cedar), Chamaecyparis obtusa (Japanese cypress), Pinus densiflora (Japanese red pine), and broadleaf trees.