Deep Learning for Improved Individual Tree Detection from Lidar Data

The aim of this study is to assess the benefits of deep learning (DL) for individual tree detection from high-density airborne LiDAR data in a complex Alpine forest ecosystem. We transformed the point cloud into 2.5D data sets and used several data augmentation procedures to deal with the sparse reference data for YoloR and ScaledYolo deep neural networks (DNN). We found that the correct detection rate is up to 15% higher for ScaledYolo compared to YoloR, but at the cost of a higher commission error. ScaledYolo outperforms traditional approaches by 20% higher detection rate. However, future research needs to deal with commission errors and to better separate the effect of sparse reference data from the intrinsic DNN accuracy.