The use of UAV derived photogrammetry data as a tool for tree measurement in industrial plantation forestry with the focus on different Local maxima and Machine learning detection methods.

Date
2023-03
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
ENGLISH ABSTRACT: With the growing interest in precision forestry, the use of low-cost multi-rotor UAVs in combination with photogrammetric software as a remote sensing tool in forestry has increased. Conventional forest inventory measurement is laborious and time-consuming which results in high operational costs, thus driving interest in alternative ways of obtaining accurate forest information. This study aimed to test the efficacy of multi-rotor UAVs structure from motion point clouds to be used as reliable source of data in forest mensuration practice, particularly in forest inventories with the focus on stocking and tree height. The study included three independent study sites planted with three different species at different planting densities which were observed at three different operational time frames. These three timeframes were before first thinning, after first thinning and before second thinning creating datasets with different conditions in terms of spacing and ground visibility. During this study three different DTM creation and normalization methods were considered, two structure from motion interpolated DTMs namely: UTG – which only consisted of compartment roads with interpolated surface between roads; and TG – which consisted of after first thinning visible ground points and compartment roads interpolated. The third DTM creation method was a DTM obtained from LiDAR data. The study also included three different tree detection methods the first two methods being local maxima based namely: CHM raster-based method where detections were done on the canopy height model and PTC Point-based method where detection were performed on the point cloud itself. The third tree detection method used was the machine learning approach where a custom object detector was trained to detect tree crowns from the created orthomosaic. The findings confirm that UAV Structure from Motion data is a valuable source of forest information. Both local maxima detection algorithms proved to be moderately reliable methods to detect trees and extract accurate information. However, optimal performance was found at lower tree density and increased levels of ground visibility. The Structure from Motion DTM created from images taken after first thinning, could produce accurate results comparable to LiDAR-derived DTM methods and manual infield statistics. Accurate stocking and tree height statistics were obtained, with mean average error ranging from -12.3% to 1% for stocking and 0.95% to 1.25% for mean tree height and 3.2% to 3.3% for dominant height. Therefore, it was shown that SfM derived DTMs obtained from after first thinning data capture flights, are capable of effectively normalising point clouds and providing accurate inventory statistics. Lastly, a few recommendations and findings from this study are mainly with regards to the capture of manual infield data; special care should be taken to minimise any measurement biases when the DBH height pairs are selected. It was also found that even though the machine learning detection method in this study did not produce reliable results, it could prove to be useful in detection of trees at lower planting densities and that it poses great potential as species classification tool as well.
AFRIKAANSE OPSOMMING: Met toenemende belangstelling in presisie bosbou, neem die gebruik van laekoste multi-rotor UAV's in kombinasie met fotogrammetriese sagteware as afstandswaarnemingsinstrument in bosbou toe. Konvensionele bosinventaris meting is moeisaam en tydrowend wat lei tot hoë bedryfskoste en dus die toenemende belangstelling in alternatiewe maniere om akkurate pantasie inligting te bekom. Die doel van hierdie studie was om die doeltreffendheid van multi-rotor-UAV's en die gebruik van fotogrametirese 3D punt wolke te toets om as betroubare bron van inligiting in bosmetingpraktyke, veral in plantasie inventarise met die fokus op plantasie digtheid en boomhoogte mates. Hierdie studie het drie onafhanklike studieareas ingesluit wat geplant was met drie verskillende spesies teen verskillende aanfaklike plantdigthede, Die data insameling het drie verskillende operasionele tydraamwerke ingesluit. Hierdie drie tydraamwerke was voor eerste verdunning, na eerste verdunning en voor tweede verdunning wat datastelle met verskillende toestande in terme van spasiëring en grondsigbaarheid ingesluit het. Tydens hierdie studie was drie verskillende DTM-skeppings- en normaliseringsmetodes oorweeg, twee fotogrametriese DTM's naamlik: UTG – wat slegs bestaan het uit plantasie paaie met geïnterpoleerde oppervlak tussen die paaie; en TG – wat bestaan het uit die na 1ste verdunning grondpunte sigbaar en plantasie paaie geïnterpoleer. Die derde DTM-skeppingsmetode was 'n DTM verkry uit LiDAR-data. Die studie het ook drie verskillende boom identifiseeringsmetodes ingesluit, waarvan die eerste twee metodes op lokale maksima gebaseer is, naamlik: CHM rastergebaseerde metode waar identifiseering op die blaredak hoogtemodel gedoen is en PTC Puntgebaseerde metode waar identifiseering op die 3D punt wolk self uitgevoer word. Die Derde identifiseeringsmetode wat gebruik was, was die masjienleerbenadering waar 'n pasgemaakte voorwerpclasifiseerder opgelei word om boomkrone op die geskepte ortomosaic te identifieseer. Die bevindinge volg dat UAV-fotogramartie 'n waardevolle bron van plantasie inligting is. Beide lokale maksima identifiseeringsalgoritmes was matige betroubare metodes om bome te identifiseer en akkurate inligting te verkry. Met optimale werkverrigting teen laer stamdigtheid en toeneemende grondsigbaarheid. Fotogramatire DTM wat geskep is van data na die eerste verdunnings data, kan akkurate resultate lewer wat vergelykbaar is met LiDAR afgeleide DTM metodes en konvensionele grond meetings statistieke. Akkurate plantings digtheid en boomhoogte statistieke was verkry, met gemiddelde fout wat wissel van -12.3% tot 1% vir plant digtheid en 0.95% tot 1.25% vir gemiddelde boomhoogte en 3.2% tot 3.3% vir dominante boom hoogte. Dus, bewys dit dat fotogrametrie afgeleide DTM's verkry van opnames na 1ste uitdunning in staat is om punt wolke effektief te kan normaliseer en akkurate voorraadstatistieke te verskaf. Laastensr n paar aanbevelings en bevindinge uit ons studie hoofsaaklik in verband met die vaslegging van in veld invemtaris metings data aangesien spesiale sorg geneem moet word om enige metingsvooroordele te minimaliseer wanneer die DBH-hoogtepare gekies word. Ons het ook gevind dat alhoewel ons Masjienleer identifiseeringsmetode nie betroubare resultate opgelewer het in ons geval nie,dat die metode steeds nuttig kan wees in die identifiseering van bome teen laer plantings digthede en dat dit ook groot moonlike gebruike bewys as spesie klassifiseerings instrument.
Description
Thesis (MSc)--Stellenbosch University, 2023.
Keywords
Citation