Fusing aerial photographs and airborne LiDAR data to improve the accuracy of detecting individual trees in urban and peri-urban areas

Abstract

Urban trees provide essential social, economic, and environmental benefits. The sustainable management of urban trees often requires basic information at the individual tree level. Aerial photographs and airborne LiDAR are two primary remote sensing data sources widely used in developed countries for large-scale mapping of individual trees in urban areas. However, limited by the imaging principles of different data modes, achieving high mapping accuracy for individual trees using either of these two datasets alone is challenging. In this study, we aimed to leverage the respective advantages of aerial photographs and airborne LiDAR to improve the detection accuracy of individual trees. Utilizing a RetinaNet-based deep learning model, we first identified key metrics from aerial photographs and airborne LiDAR data for individual tree detection. Then, we rectified the misalignment of individual trees between the aerial photographs and airborne LiDAR data using a newly described object-oriented approach. Finally, we detected individual trees at the pixel level and the decision level, respectively. For pixel-level fusion, we combined the selected metrics (i.e., the red, green, and infrared bands as well as the canopy maximum model) from two datasets to detect individual trees. At the decision level, we fused the crowns of individual trees detected from the two rectified datasets. Our findings reveal that rectifying the misalignment between individual trees in both datasets significantly enhances detection accuracy, resulting in a notable increase in F1-score from 0.724 to 0.828. Furthermore, our results indicate that the decision-level data fusion approach yields the highest detection accuracy, with an F1-score of 0.814. This performance surpasses that of aerial photographs (F1-score: 0.592) and airborne LiDAR (F1-score: 0.776) individually. Our study underscores that integrating aerial photographs and airborne LiDAR data is an effective approach to improve the detection accuracy of individual trees in heterogeneous urban and peri-urban landscapes.