Development of an Urban Vegetation Monitoring System using Image Processing and Deep Learning

Automated monitoring of urban vegetation, particularly trees, facilitates large-scale urban planning and environmental surveillance. However, such systems can be cost-prohibitive to deploy, especially in smaller cities. This paper describes a cost-efficient standalone RGB-imaging platform with adaptable mounting for service vehicles, implemented in Hamilton East, Hamilton City, New Zealand. First, the hardware design and imaging platform are described. Then, to evaluate the quality of the tree imagery captured by this system, two deep-learning models, YOLOv5n and YOLOv8n, were trained using the collected images. The collected dataset consists of approximately 12,000 images of various tree species typical to the region. Of these, 450 images (4%) were used for training and validation, and the remaining images were used for testing. The models were then tested on images collected in 4 different test image collection sessions. The results of the models in the validation phase showed that the YOLOv5n model demonstrated a precision of 0.87 and a recall of 0.85, compared to YOLOv8n’s precision of 0.86 and recall of 0.82. During the test phase, both models performed comparably, indicating that the imaging system can capture high-quality images before direct use with deep-learning algorithms for tree detection. Furthermore, the imaging system is capable of capturing at least 11 images per tree on an urban route, showing the potential to be deployed for urban vegetation monitoring.