Real-time Urban Forest Management Using Machine Learning

Abstract

In the context of a more sustainable world and smart cities solutions, the environment surveillance arises as a relevant topic of research to reduce environmental pollution and optimize the use of natural resources. Real-time urban monitoring and analysis of tree mortality, for instance, are essential for evaluating possible deforestation or degeneration of urban green areas.This research aims to develop a real-time urban forest management tool to support managers' decision making and consequently allow cities to offer better quality of life to the population. The proposal fits the context of smart and sustainable cities with the focus of supplying the lack of solutions for cities that do not have trees and green areas in the proper urban conditions. Therefore, the topic of research proposed in this project is essential for the proper management of these urban areas, particularly for reducing the risk of tree falling and making improvements to the city and population by taking advantages of the functions of green areas and forest fragments (such as flood reduction, wildlife preservation and reduction of air pollutant levels). In order to attain the above-mentioned benefits of the green area surveillance, algorithms for image processing and machine learning techniques, specifically the ones related to deep learning models, are expected to be used to extract the main features of the images associated with trees having the risk of falling in the urban environment. Images acquired from Unmanned Aerial Vehicles or portable cameras will be used for subsequent analysis by the proposed algorithms. A real-time approach for monitoring green areas in the urban environment becomes necessary owing the time-consuming diagnostic methods currently used for urban forest and green area management. Therefore, the proposed research will contribute for automating the current situation of the urban region by means of a real-time image analysis of the target areas.This project will make the use of supervised machine learning techniques by paying special attention to several models known as Deep Learning, which represent the state of-the-art in image processing and analysis. Convolutional Neural Networks (CNN) will be used for segmentation and identification of the objects of interest. Since the intention is the usage of a supervised machine learning-based approach, a dataset of images where the regions of interest are manually annotated will be created for building the CNN models. Once the target features are defined and a training dataset is created, CNN models will be trained to detect the defined features.In terms of CNN architectures, several models will be evaluated in order to define an optimized and accurate architecture for the identification of the objects of interest. Specifically, the U-Net, Faster R-CNN and You Only Look Once (YOLO) architectures will be evaluated to identify the candidate regions that will be latter classified. An important distinction must be made between the tasks of object detection and semantic segmentation. In the semantic segmentation, each element of the image (i.e. pixel) has its own class, whereas in object detection the class is determined by a set of pixels. For example, if the target visual feature is "abnormal coloring leaf", a CNN model detects each leaf separately, while a segmentation network will classify each pixel of the image into "abnormal coloring leaf" or "background/other region". The choice of the architectures will be made based on the volume of data, available hardware, and desired accuracy. For the segmentation task, we intend to evaluate the U-Net architecture for the identification of objects based on the semantic segmentation, while the Faster R-CNN and YOLO architectures will be used for instance detection. Traditional architectures used for classification tasks, such as Alex Net and VGGNet, will also be evaluated for classifying the detected objects.