Semantic Classification of Urban Trees Using Very High Resolution Satellite Imagery

There is an urgent need for urban tree classification, in order to assist with ecological environment protection and provide sustainable development guidance for urban planners. While most of the existing studies have concentrated on tree crown extraction or tree species identification, only a few studies have attempted to conduct semantic classification of urban trees from an urban habitat perspective. The lack of semantic information means that it is difficult to meet the needs of ecological and environmental issues. As such, in this study, a novel three-level (pixel-object-patch) framework for semantic classification of urban trees is proposed to categorize urban trees as park, roadside, and residential-institutional trees. These three categories are cognized and conceptualized by humans and serve as different ecological functions in urban areas. Park is important urban greenery accommodated within recreational and cultural facilities. Roadside and residential-institutional trees are distributed along streets or in neighborhoods. The framework for the semantic classification of urban trees includes the following steps: 1) vegetation information extraction at the pixel level utilizing a spectral vegetation index; 2) vegetation-type classification at the object level employing spectral and textural features; and 3) urban tree classification at the patch level, where a series of metrics related to area, shape, structure, and spatial relationship are considered. Two typical Chinese megacities, Shenzhen and Wuhan, were chosen to demonstrate the applicability and effectiveness of the proposed method. The results reveal that the proposed method can achieve a satisfactory performance, with the overall accuracy reaching 85%. Moreover, the producer's and user's accuracies are generally high for most tree categories (>80%). The further landscape analysis demonstrates some general characteristics of the natural landscape configuration: residential-institutional trees show greater fragmentation and spatial heterogeneity, and park trees show the maximum physical connectedness and aggregation.