Detection of Green Roofs through Satellite Imagery Using Faster R-CNN

Abstract

<p>With the threat of global warming, there is a need to implement sustainable solutions. Due to urbanization, forests and other natural resources that could have helped reduce carbon emissions are being depleted. Thus, cities, like Toronto, experience urban heat island effects, food shortages and increased carbon emissions. To help solve this, Green Roofs are encouraged to be implemented in Toronto to help reduce the mentioned issues. In addition, green roofs are difficult to maintain and COVID-19 making it more difficult for urban farmers to physically go and check the green roofs. Hence, this study proposes to use deep learning in detecting green roofs from satellite imagery. The approach uses NDVI values and Toronto Building data to obtain green roof locations. From here, satellite images were taken from map services such as Google Earth Pro and ArcGIS Pro. The images were preprocessed and fed into Faster R-CNN model to detect green roofs. Since this is a first step in creating a system that can monitor the Green Roof Food-Energy-Water nexus, in the future, the behaviour of Faster R-CNN with Toronto Roof Satellite Images was observed. It was shown that the Faster R-CNN provided a higher validation accuracy and higher confidence score than the Fast R-CNN. Additionally, it is suggested that the current Faster R-CNN model needs to be improved, to precisely detect green roofs in different image resolutions and dimensions.</p>