Urban street tree species identification and factor interpretation model based on natural images

Urban street trees bring the beautiful ecological environment for human beings, but also may harm human health. Tree pollen is an important allergen that causes people to suffer from asthma and rhinitis, causing a serious medical burden. In order to protect human health and reduce medical costs, urban street trees need to be accurately identified. However, the identification of the urban street tree is influenced by the natural image (images with light intensity, season and shooting conditions), tree characteristics and identification models. To solve the problem, we proposed an interpretation model for identifying tree from natural images, named Ev2S_SHAP. Then, we applied the method to explain the influence of environmental and leaf factors on tree recognition and the identification accuracy. The open-source natural image dataset of urban street tree with complex situations such as different angles, distances, light, and times was used as the research object to verify the proposed tree identification model. The results showed that the overall evaluation of the identification results for 50 tree species: recall, precision, accuracy, and F1 score were 98.12 %, 98.18 %, 98.11 %, and 98.13 %, respectively. The identification accuracy of deciduous shrubs, evergreen shrubs, deciduous trees, and evergreen trees was 98.85 %, 98.47 %, 98.76 %, and 96.38 %, respectively. Complex lights and angles, long-range shooting, and winter conditions weakened the extraction of leaf features and were not conducive to the identification of tree species. Leaf shape characteristics had important influence on tree species identification. The effects of circularity, minimum circumcircle, and perimeter on tree identification accounted for 94.36 %, 75.61 %, and 69.12 %, respectively. Circularity was positively correlated to the identification contribution of elliptic leaf species, but opposite to that of lanceolate leaf species. The perimeter contributed positively correlated to the identification of lanceolate leaf species, but the minimum circumcircle was negatively. The Ev2S-SHAP effectively improved the identification accuracy of tree species. The study provides an innovative method for identifying and interpreting tree species.