Urban tree crown detection based on deep learning and high-resolution aerial imagery: PTCNet for Pullman, WA, USA

Individual tree data in urban settings are used for many purposes, and gathering such information requires time and other limited resources. Additionally, the data collected are spatially and temporally sparse, especially for continuous monitoring. However, high-resolution images and deep learning can offer automated and accurate detection of trees in complex urban settings. Therefore, this study compared four popular convolutional neural network CNN-based object detection models (You Only Look Once v3, RetinanNet, Mask R-CNN, and Faster R-CNN) to map individual trees. We used high-resolution aerial imagery (∼8 cm spatial resolution), which was manually annotated to derive training (4,859) and testing (1,184) datasets. The analysis was carried out in three phases: First, we trained all the models for 20 epochs and evaluated the performance using standard metrics (Precision, Recall, and F1 score). Second, the best model was selected and retrained longer (30 epochs) with more data (5002 annotations) to develop an urban tree crown detection model for Pullman – a small-sized city in the inland northwest of the United States. Finally, we tested the reliability of the developed model under two scenarios. According to our analysis, YOLOv3 (F1 score: 69 %) outperformed Mask R-CNN (F1 score: 60 %), RetinaNet (F1 score: 57 %), and Faster R-CNN (F1 score: 52 %). Based on the evaluation metrics and visual assessment, YOLOv3 was selected to develop the final urban tree crown detector – Pullman Tree Crown Network (PTCNet), for our study area. PTCNet had precision and recall values of 78 % and 62 %, respectively. It also performed well under different tree arrangements, achieving an F1 score of over 70 %. The model was used to generate ∼12,000 individual tree locations. Subsequently, height information was extracted from a LiDAR-derived canopy height model, and a comprehensive tree inventory dataset was derived. The model and dataset are publicly available ( https://github.com/Okikiola-Michael/PTCNet ) for different applications, thus, contributing to open science. This study provides a straightforward and repeatable framework for researchers and managers to map urban trees with height information, which is useful for spatial and temporal tree monitoring. This study further highlights the performance of four popular models and supports the application of deep learning and aerial imagery for individual tree detection in complex urban settings. • Compared the performance of four state-of-the-art object detection models. • YOLOv3 performed better than RetinaNet, Faster R-CNN, and Mask R-CNN. • Developed the first urban tree crown detector for a city in the inland northwest of the USA. • Developed the first publicly available database for the city with individual tree crowns and height information. • Provided relevant data that supports urban tree monitoring and management.