University of Twente Student Theses


Detection of Japanese Knotweed Beside the Road using Deep Learning

Surya, Siddharth (2023) Detection of Japanese Knotweed Beside the Road using Deep Learning.

[img] PDF
Abstract:Invasive Alien Plant Species (Japanese Knotweed) are a threat to biodiversity. Its monitoring can help conserve the native flora and fauna. Japanese Knotweed habitat is generally along the roadways and, its identification, can prevent its spread. The research discusses the detection of the invasive Japanese Knotweed plant using a mobile phone camera mounted on a bicycle. It mainly focuses on object detection models under two categories: (1) Generic and (2) Few-shot. Generic object detection has been used to detect invasive plants. However, a lack of pre-labeled datasets makes this approach expensive and time-consuming. Moreover, the object detection model considering the plant phenological cycle further increases the requirement for annotated images. Thus, the few-shot object detection algorithm offers an alternative approach with limited annotated images. However, there is no previous study on its performance for invasive plant species (Japanese Knotweed). Thus, this study aims to bridge these gaps. It considers the creation of a realistic dataset for detecting invasive Japanese Knotweed plants using a mobile phone camera mounted on a bicycle. A performance comparison between the extensively used general object detection i.e, Faster RCNN, state-of-art YOLOv7, and few-shot state-of-art ‘DeFRCN’ is performed under PASCAL VOC settings of 1, 2, 3, 5, 10 images and complete dataset. Resolution & support set study concerning DeFRCN is also discussed. It was found that under limited training images (1, 2, 3, 5, 10), the generic object detection overfits and few-shot object detection model with data augmentation offers 5.33, 4.83, 6.45, 7.31, 8.30 mAP50 on test set respectively. YOLOv7 offers the highest mAP50 test with a complete dataset, which is 32.6. On the contrary, the Faster RCNN has a large false positive and is not robust. Overall, the research focuses on contributing to the conservation of native flora and fauna.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Embedded Systems MSc (60331)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page