Computer Vision-Based Roof Detection using YOLOv4

Over the last few decades, there has been a shift of focus in urban planning in which artificial intelligence (AI) is being integrated to make informed decisions and predictions through modelling and simulation. Among all AI applications, computer vision (CV) plays a crucial task because of the role of visual data in the field of urban planning. Application of CV like object detection specifically roof detection has been very useful for the data collection and interpretation process of urban planning. At the top of this technology is the YOLO model which is an object detection model optimized to accurately detect objects based on the dataset trained within it. This research utilized a YOLOv4 model to detect roofs based on Google Earth satellite images. A total of 500 ground truth images were extracted across different regions of the Philippines and were annotated using Computer Vision Annotation Tool (CVAT) which were then pre-processed and augmented resulting to a total of 4,802 images. Using a custom dataset, the YOLOv4 model demonstrated superior performance with a mean average precision (mAP) of 92%, outperforming alternatives such as Single Shot Detection (SSD) MobileNet and Detectron 2 Faster R-CNN, which achieved maximum mAP values of 83% and 79%, respectively. This highlights YOLOv4's accuracy in roof detection from Google Earth satellite images and proving its potential in urban planning applications.