Estimating Forest Variables from LiDAR Pointcloud Data Using a Deep Learning Approach

Abstract: Knowledge about forest measurements is essential for efficient and sustainableforestry. One important measurement is wood volume, both from an economicand an environmental perspective. Therefore it is crucial to measure wood vol-ume accurately and reliably. With airborne laser scanners, wood volume can beestimated at a large scale, more time efficiently than traditional, manual measure-ments. By utilising deep neural networks, we present methods to predict woodvolume on point clouds efficiently and accurately. Different network structuresfor point cloud regression are devised using field measurements from wood har-vesters and manual field measurements. To achieve more data, techniques tosplit up areas into smaller subareas and data augmentation methods were imple-mented. Our version of GDANet adapted for regression provided the best resultswith a RMSE of 62.68, MAPE of 24.7%, and relative RMSE of 28.0%. Furthermore,the final model produces more accurate wood volume predictions than more shal-low machine learning methods and predictions from Skoglig Grunddata.