Assessing the Vitality of Urban Trees using Remote Sensing and Deep Learning

Abstract

Abstract The use of Convolutional Neural Networks (CNN) has been widely implemented in forestry-related tasks as species classification, crown detection and mortality identification. The usage of several sources as images, point clouds and elevation models have generated relevant results in different forested areas, but unfortunately these studies have not been focused on urban trees. Therefore, the objective of this study is to investigate the performance of CNN for classifying the vitality of urban trees, which are increasingly affected and stressed by the Urban Heat Island Effect. Aerial and Sentinel-2 images are sampled for feeding the CNN model. The prediction of the vitality classes shows a precision of 74,69%, especially for the most represented class (healthy trees). The achieved results allow to better understand the performance of a CNN network for determining the vitality of trees in an urban context where diversity of vegetation patterns can represent a big challenge for classification tasks.