3D Semantic Segmentation of Airborne & UAV LiDAR Point Clouds for Precise Tree Counting: a Machine Learning Approach

In a world where sustainable forest management and understanding of our ecosystems have become priorities, accurately and efficiently counting trees, especially in geographically challenging areas, has become a major challenge. This article explores an innovative approach that combines the power of machine learning and advanced technologies, such as airborne and drone lidar, to address this critical challenge. After data collection and preprocessing, we show how semantic segmentation using the "3DMASC" method makes it possible to distinguish the different classes (ground, high and medium vegetation) with remarkable accuracy. We then quantify the number of individual trees by applying, for the first time, the "Treeiso" clustering method, traditionally reserved for terrestrial data, to data acquired by airborne sensors. The comparative analysis we performed provides an unequivocal answer. Using two different datasets, one acquired by an airborne LiDAR Galaxy T1000 and the other by a Zenmuse L2 drone LiDAR, we show that the qualitative results of semantic segmentation and tree isolation are remarkably close to the ground truth. The quantitative results show significant accuracy for both datasets. In fact, for tree isolation, we achieve an error rate of less than 10% for both data types. Furthermore, the overall accuracy of the two semantic segmentation models is around 0.92.