2D-CNN for tree stem detection and segmentation using data collected with a harvester-mounted mobile laser scanner

Forestry operations planning has typically relied on forest inventories, which require manual field measurements, a process that is often costly in terms of money and labor. Advancements in remote sensing technologies have demonstrated the potential to enhance both the accuracy and efficiency of forestry operations planning, thereby reducing costs and labor requirements. Integrating these technologies with forestry machinery necessitates automation, particularly in harvesters. The need for an exact three-dimensional description of tree stems has become a critical requirement. However, stem detection and segmentation from LiDAR data have typically required extensive post-processing, which limits their applicability in scenarios where real-time information is critical. This work examines the feasibility of detecting and segmenting tree stems in real-time using LiDAR data obtained through mobile laser scanning (MLS), creating a proof of concept. To achieve this, each frame captured by the sensor is processed independently and transformed into a 2D projection, where a Convolutional Neural Network (CNN) identifies and segments tree stems. This approach eliminates the need for extensive point cloud analysis, enabling faster processing and rapid response in operational environments. Detecting stems through this method offers distinct advantages over point-based methods. To fully validate this proof of concept, further development and integration of the algorithms will be necessary, along with expanding the training dataset to improve model performance. Accurate real-time detection and segmentation of tree stems and logs could be the basis for the automation and robotization of forestry operations, enhancing efficiency and precision in forest management.