Enhanced Solar Potential Analysis: Separating Terraced House Rooftops Using Convolutional Neural Networks

Solar power, a clean and renewable energy source, plays a pivotal role in achieving sustainable development goals by offering affordable, reliable, modern energy solutions and mitigating energy-related emissions and pollutants. Current studies predominantly focus on solar potential analysis derived from machine learning-based rooftop area segmentation. However, these studies reveal an overestimation of usable area for solar output calculations in terraced houses, due to failing to distinguish individual households within terraced structures. This research delineates state-of-the-art Machine Learning and computer vision techniques applied on remote-sensing images obtained via the Google API. The dataset, manually annotated 300 training images and augmented to 1500 training images and 150 validation images, is focused on the UK, particularly terraced house areas. The stand-alone Convolutional Neural Network used to segment terraced-structure rooftop areas reaches an intersection over union of 82.34%. The model uniquely addresses the segmentation of contiguous terraced houses in the UK, which is pivotal for the solar installation assessments in the UK's residential landscape.