Vegetation detection using a hybrid method of vegetation indices and convolutional neural network

Thumbnail Image
Lim Soon Eng, Mr.
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Journal Issue
Vegetation inspection and monitoring is time-consuming. Unmanned aerial vehicle (UAV) or drone can be used for the tasks but most drones has limited spectral bands (such as RGB camera) which restricts advanced vegetation analysis. Additional spectral bands can produce more accurate analysis but are costly. Vegetation indices (VI) is a technique to maximize detection sensitivity related to vegetation characteristics while minimizing others. Machine learning (ML) may also improve detection accuracy in vegetation analysis. This study explores VI techniques in identifying vegetation objects including hybrid VI and ML to overcome the limitation of existing VI techniques. The hybrid methods were analysed and evaluated to find the strengths and limitations to improve detection accuracy. Several VI methods such as Visible Atmospheric Resistant Index (VARI), Green Leaf Index (GLI), and Vegetation Index Green (VIgreen) were combined with the ML technique; You Only Look Once (YOLO). Data for testing were collected from aerial images of a number of locations. Hybrid segmentation, a process to divide the targeted pixels and then eliminate the non-vegetation pixel in the image were performed. Selected VI techniques were applied on several objects of the same images with varied performance. The performance of hybrid methods developed in this study was measured by the accuracy of vegetation detection. The results obtained showed that >70% of the vegetation objects in the images were accurately detected. The hybrid segmentation has generally increased the accuracy compared to the initial hybrid method. Mixture of VARI and YOLO in hybrid segmentation method performs best at 84% detection accuracy. GLI and YOLO on the other hand, gave 81% detection accuracy and VIgreen with YOLO gave 78% detection accuracy. There are several limitations with the proposed methods identified throughout the experiments. GLI + YOLO combination is less sensitive in detecting tiny tree and occasionally misdetect tree shadow as vegetation. Despite the limitations, hybrid segmentation shows an improvement in reducing misdetection of greenfield as vegetation compared to the initial hybrid method. Overall, the proposed hybrid segmentation method and hybrid detection method have the ability to better detect vegetation objects from aerial images data, as compared to the VI techniques alone.
Vegetation detection