Urban tree canopy mapping -an open source deep learning approach

Abstract: Urban trees have an important role to provide ecosystem services and to make our cities greener and more sustainable. The changing climate and densification of cities make it even more valuable to preserve and investigate in urban trees. Tree canopy detection in cities is challenging, with both trees and other objects of irregular shape, size and complexity. The aerial images from Lantmäteriet (the Swedish mapping cadastral and land registration authority) is a great source for image analysis on high resolution images, but the leaf-off images may be challenging for tree canopy detection. The aim of the study was to test if Light Detection and Ranging (LIDAR) could be combined with aerial images to develop a deep learning tool for urban tree canopy mapping under leaf-off conditions. The deep learning method using LIDAR and aerial leaf-off images had a precision of 88 % mapping urban tree canopy. Using the same method with LIDAR and Infrared (IR) leaf-on data yielded a precision of 91 %. The LIDAR data increases the accuracy for leaf-off data when added to the deep learning model. The findings of this study indicate that tree canopy mapping with LIDAR and aerial images taken under leaf-off conditions can be used for tree canopy mapping with comparable results to other methods.