An investigation of three machine learning models using UAV derived photogrammetry for detecting and measuring tree stumps

Abstract

Accurately quantifying stump volume on post-harvested sites is required to assess potential volume gains for biomass utilisation. The relatively uniform distribution and shape of stumps across such sites makes them well-suited for detection using machine learning (ML) algorithms. Recent developments in the analysis of Digital Aerial Photogrammetry (DAP) data acquired by unmanned aerial vehicles (UAVs) have enabled the reliable identification of stumps via advanced ML methods. Furthermore, the processed outputs from these algorithms provide estimates of stump diameter and height, facilitating calculations of biomass volume. This integration of UAV-based photogrammetry and ML techniques presents a promising approach for enhancing forest management and biomass assessment. In this study, we trained three different ML model types: Faster Region-based Convolutional Neural Network (R-CNN), Single Shot Multibox Detector (SSD) and You-Only-Look-Once (YOLO). The data for the virtual stump detections came from two Norwegian sites, with stumps of Picea abies (L.) H.Karst., and three South African sites, with stumps of Pinus patula Schiede ex Schltdl. & Cham. We assessed the detection rates of each model and compared metrics by using similarly annotated images. The resultant encapsulating bounding boxes of detected stumps were used to calculate diameters and compared to field measurements. Each bounding box is rectangular in shape, and the average of the height and width was calculated to get an estimated diameter value. Virtual stump heights were determined from the Digital Surface Model (DSM) by subtracting the mean height of the surrounding area from the mean height of the stump. The calculated heights of the stumps can be used to assess potential loss of wood volume due to inefficient harvesting techniques. Similarly, the calculated wood volume can be used to estimate residual biomass, and therefore assist Foresters in deciding how best to utilise these stumps. Visible stumps on post-harvested sites could be detected with high rates of accuracy, with almost perfect precision from some object detection models, albeit at low levels of recall. Overall, all three model types had an F1-score of above 73% with the best model attaining an F1-score of 89%. Stump diameters were generally overestimated and this was not found to be related to stump size. Stump heights were underestimated in most cases.