Tree-Net: A novel deep learning tree detection architecture using UAV LiDAR data

Abstract

Individual Tree Detection (ITD) is essential for assessing tree parameters including spatial distribution, geometrical characteristics, and species identification within forested landscapes. This process is vital for tree inventory management and forest carbon accounting. Such information on individual trees can be derived from high-resolution Unmanned Aerial Vehicle (UAV) Light Detection and Ranging (LiDAR) data by utilizing Deep Learning (DL) object detectors. Unlike traditional RGB images, UAV LiDAR data can be rasterized into multiple bands (or features) representing different structural and textural attributes, such as Vertical Distance (VD), Canopy Height Model (CHM), and texture-based features, to provide richer information to enhance detection. However, existing methods for UAV LiDAR data face challenges in processing high-resolution data due to complex forest environments, overlapping tree crowns, and limitations in utilizing rich multi-band information (i.e., input raster layers with more than three bands). To address these issues, this paper introduces Tree-Net, an innovative one-stage DL structure object detection for ITD on rasterized UAV LiDAR data. Inspired by a deep neural network based on You Only Look Once (YOLO), Tree-Net attempts to benefit from greater numbers of input bands to obtain sophisticated spectral and spatial features. By optimizing the neural network structure, Tree-Net improves computational efficiency and detection accuracy by employing a shallower architecture. Tree-Net's performance is evaluated on two distinct datasets, Heiberg and ForInstance benchmarks consisting of around 25,000 trees including deciduous and coniferous trees to represent different canopy shapes, heights, and densities. The Best detection metrics results indicate that the proposed method outperforms existing literature (i.e., YOLOv5) and Modified YOLO by achieving 7 %, 41 %, 30 %, and 34 % improvement in accuracy, precision, recall, and F1-score metrics, respectively. In terms of time efficiency, Tree-Net outperformed our previously developed modified version of YOLO (Modified YOLO), which accepts multi-band instead of classic three-band (RGB) input, by 60 % on both training and testing phases. From a practical perspective, Tree-Net holds significant promise for urban tree inventory, forest management, and satellite ground-truth data generation for large-scale applications including biomass estimations and carbon exchange.

Topics & Concepts