Counting trees in a subtropical mega city using the instance segmentation method

Abstract

Counting the trees in a given area or city is meaningful when making decisions for government policies and administration, including the international afforestation effort (i.e., the Trillion Trees Campaign). Determining individual trees on a large scale poses significant challenges, especially in subtropical and tropical areas, because of their diverse crown characteristics. Using very high-spatial-resolution images, we can see the tree crowns clearly. In this study, we counted the population number of trees in the subtropical mega city of Guangzhou, and we used an end-to-end tree-counting deep-learning framework in the regional-scale tree detection by delineating each tree crown. It is a simple framework in which individual trees can be detected directly without manual operation. We used the cascade mask regions with convolutional neural networks (CMask R-CNN) as the backbone and added three types of attention modules to build the derivatives of the CMask R-CNN. The experimental results showed that the CMask R-CNN performed the best among all of the methods, and more than 112 million individual trees with crown sizes of large than 1 m2 were detected. The experimental assessment indicated that the accuracy was 88.32% in terms of the R2 value and 82.56% in terms of the F1-score. This study not only revealed the number of trees, but it also provided the tree density at different scales, which is a prominent component of the ecosystem structure. We also analyzed the tree density at the 30 m and 1000 m scales. The experimental results showed that Guangzhou has a high canopy cover with a tree density of 150 trees per hectare. For the entire city, the tree density is highest in the northern area of Guangzhou, followed by the central part. From the central-western part to the central-eastern part, the tree density increases. The lowest tree density is located in the southern part of Guangzhou near the Pearl River estuary, which is a basic farmland conservation area. It is notable that the tree density of the urban land is about 15 trees per 900 m2, indicating a good living environment. The method developed in this study provides a flexible means of large-scale tree counting without manual operation based on very high-spatial-resolution images.