Advancing tree detection in forest environments: A deep learning object detector approach with UAV LiDAR data

The initial phase in determining tree parameters within urban and forest environments is Individual Tree Detection (ITD), which includes tree count, spatial distribution, height, volume, crown dimensions, and species identification. This process holds significance in applications like urban forest inventory, planning, and tree carbon accounting. Traditional airborne and spaceborne remote sensing data lack the precision required for ITD due to their coarse spatial resolution. High-resolution multispectral and Light Detection and Ranging (LiDAR) data collected by Unmanned Aerial Vehicle (UAV) sensors have enabled detailed ITD and tree parameter estimation. However, the processing of such very high-resolution data presents challenges. While existing algorithms for processing 2-dimensional (2D) and 3-dimensional (3D) data from airborne sensors exist, they prove impractical for UAV data, primarily due to its extremely high spatial resolution. Recent strides in deep-learning algorithms offer promising solutions for ITD using UAV data. This paper introduces a novel ITD method using a modified You Only Look Once V7 (YOLO V7) deep learning object detection framework, employing UAV LiDAR data. The approach involves rasterizing point clouds in various channels, including Vertical Density (VD), Canopy Height Model (CHM), Gradient of the CHM (G-CHM), and Local Binary Pattern of the CHM (LBP-CHM). Subsequently, the YOLO V7 object detector is employed to identify the bounding box of each tree. The modified YOLO7 algorithm is trained and tested on UAV LiDAR data collected over diverse regions of interest, encompassing pine, deciduous, and mixed tree types with varying tree densities. The results exhibit a substantial enhancement over the previously developed YOLO3 on airborne LiDAR data, showcasing heightened accuracy, precision, and recall within the ranges of 0.7-0.94, 0.76-0.99, and 0.8-0.97, respectively. From a practical standpoint, our automated method holds potential for urban tree inventory updates and serves as a valuable tool for ground-truthing large-scale satellite-based forest structure and biomass estimation, among various other applications.