Urban Tree Detection from Remote Sensing Data based on DeepForest Model

Abstract

Abstract. The distribution and quantity of trees within a city impacts issues such as urban heat islands, air quality and general city planning. Having an automatic procedure for cataloguing them can be a valuable aid for future urban planning and design. In this paper, the forestry surveying neural network, DeepForest, is utilised for tree detection in the urban environment. The study area covers District Lozenets, which is the greenest part of Sofia, Bulgaria. Three distinct approaches are implemented considering the urban vegetation context - a simple tree detection, a tree cluster detection and a mixing approach between the two based on approximation with Poisson Disk Sampling. The evaluation of the developed models, in terms of F-1 score, shows that the achieved results are comparable to the ones achieved by the original application of the DeepForest model. Due to the specifics of urban data, all models tended to achieve a higher precision but a lower recall than the original DeepForest. Conditions, such as shade from the sun, buildings or other trees, make the detection more challenging. The obtained results prove the feasibility of the proposed approaches, even with a small amount of labelled data. The tree cluster and mixed approaches have the potential to resolve part of the issues coming from the urban environment context of application.