Seedling and juvenile tree detection using UAV photogrammetry and machine learning

Simple item page
dc.contributor.advisorTalbot, Bruceen__ZA
dc.contributor.advisorAckerman, Simonen__ZA
dc.contributor.authorKok, Simeon Jamesen__ZA
dc.contributor.otherStellenbosch University. Faculty of AgriSciences. Dept. of Forest and Wood Science.en__ZA
dc.date.accessioned2024-03-07T13:10:15Zen__ZA
dc.date.accessioned2024-05-08T11:03:26Zen__ZA
dc.date.available2024-03-07T13:10:15Zen__ZA
dc.date.available2024-05-08T11:03:26Zen__ZA
dc.date.issued2024-03en__ZA
dc.descriptionThesis (MSc)--Stellenbosch University, 2024.en__ZA
dc.description.abstractENGLISH ABSTRACT: Accurate counts of seedling survival are essential for establishing a measure of mortality rates in replanted sites. This in turn measure the effectiveness of the planting operation deployed in the stand. To further this measure, advances in digital aerial photogrammetry obtained from UAVs and machine learning algorithms (object detection models) have led to increased use of this technology in the forestry industry. Seedlings and juvenile trees are ideal objects to be detected and counted by object detection models, due to their shape and contrast in colour to background in most digital orthomosaics. In this study, three different types of object detections models were trained using annotated seedlings as its training data, to detect and count the number of seedlings in 4 replanted sites. A total of 12 models were trained, 4 of each model type (SSD, Faster R-CNN and YOLOv5). The detection rates for each model were assessed and compared to eye count data as well as data measured in the field. The orthomosaics of four different sites were used to detect the seedlings located within the site. The resulting bounding box that covers a seedling was used to determine how crown diameter (seedling size) influences detection rates. The objects causing false detections (false positives) was counted to determine which objects are most likely to cause confusion for the models. The influence of data set size on detection was also studied. The results show that seedlings can be detected with high rates of accuracy with the three types of object detection models used. The best performing model being the SSD model achieving an average recall and precision of 93% and 100% on the orthomosaics of 4 different sites. Crown diameter showed to influence detection rates, as an increase in crown diameter showed an increase in detection rates. Vegetation surrounding the seedlings, such as weeds, was the main cause for false detections for the models. The number of annotations influenced model detection rates, with more annotated seedlings increasing model performance.en_ZA
dc.description.abstractAFRIKAANSE OPSOMMING: Geen opsomming beskikbaar.en_ZA
dc.description.versionMastersen_ZA
dc.embargo.terms2024-12-31en_ZA
dc.format.extentvii, 80 pages : illustrationsen_ZA
dc.identifier.urihttps://scholar.sun.ac.za/handle/10019.1/130799en_ZA
dc.language.isoen_ZAen_ZA
dc.publisherStellenbosch : Stellenbosch Universityen_ZA
dc.rights.holderStellenbosch Universityen_ZA
dc.subjectUAVen_ZA
dc.subjectObject Detectionen_ZA
dc.subjectSeedlingen_ZA
dc.subject.lcshAerial photography in agricultureen_ZA
dc.subject.lcshMachine learningen_ZA
dc.subject.lcshSeedlingsen_ZA
dc.subject.lcshDrone aircraften_ZA
dc.subject.lcshUCTDen_ZA
dc.titleSeedling and juvenile tree detection using UAV photogrammetry and machine learningen_ZA
dc.typeThesisen_ZA
Files
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
kok_seedling_2024.pdf
Size:
5.11 MB
Format:
Adobe Portable Document Format
Description:
Download
Collections
Masters Degrees (Forest and Wood Science)