Individual Tree Crown Delineation in Heterogeneous Urban Environments Using High-Resolution Airborne Imagery

Urban trees are crucial to mitigate the increasing urban temperatures caused by climate change. Tree inventories, which are useful for urban planning and tree management, are mostly based on expensive and incomplete in-situ measurements. Freely available high-resolution aerial imagery and Convolutional Neural Networks offer promising opportunities for large-scale urban tree mapping at low costs. In this study, a Mask R-CNN model and true orthophotos of 20 cm spatial resolution were utilized for individual tree crown delineation in Berlin and Osnabrück, Germany. In Berlin, seven different urban environments including Park, Industrial, Residential and Roads & Parking areas were analyzed separately to evaluate the model's robustness. Afterwards, to test the transferability, the individual tree crown delineation model was applied to Osnabrück without using additional training data. For Berlin, the model showed good delineation performance with an F1-Score of 82.8%. Further, good delineation accuracies between 85.3% and 91.0% were achieved for six of the seven urban environments. Meanwhile, a significantly lower F1-Score of 68.6% was achieved for the remaining class Park. The model transfer to Osnabrück performed very well with an F1-Score of 84.4%. Overall, the Mask R-CNN model appears to be very suitable for individual tree crown delineation. The model can be successfully applied to entire cities with heterogenous urban environments and transferred between cities of different sizes and varying aerial imagery recording months. The delineation results can be used by urban planners for advanced investigations regarding the urban climate and as data source for further research.