Discussing street tree planning based on pedestrian volume using machine learning and computer vision

Abstract

An increasing number of cities are advocating people-oriented street tree planning to make cities more walkable, livable, and sustainable. People-oriented planning can be further divided into resident-oriented planning and pedestrian-oriented. Numerous studies and policies have started to include the distribution of residents as an essential factor while planning street trees. However, due to various reasons, existing studies failed to enhance the street tree planning based on pedestrian volume, which has a higher exposure. Our study is the first research that combines street tree and pedestrian volume in discussing existing patterns between them and proposes additional scientific planning suggestions. To achieve the research goal, we proposed a methodology framework called LightGBM with K-fold Max variance Semi-Supervised Learning and DeepLab v3+ (KMSSL-DL). KMSSL-DL combines machine learning and computer vision technology to estimate pedestrian volume with unlabeled data from high dimensional urban features, and extract tree crowns from satellite imagery in the Central Business District, City of Melbourne, Australia. KMSSL part achieved an excellent prediction effect (R2 score = 0.8360, RMSE score = 0.2304). We also used DeepLab v3+ to recognize and extract street trees from Google Earth satellite imagery with good performance (mIoU = 84.37). Lastly, we combined the two results to conduct a pattern analysis, enabling us to find four patterns between street trees and pedestrian volume: more trees - more pedestrians (MTMP), more trees - fewer pedestrians (MTFP), fewer trees - more pedestrians (FTMP), fewer trees - fewer pedestrians (FTFP). We discussed the four patterns and presented planning suggestions, including urban planners giving the most attention to FTMP.