[Dissertation] Using Machine Learning to Detect Solar Panels in Aerial Images
Masters dissertation by Palesa Rachael Lepamo, Faculty of Engineering, Built Environment and Information Technology University of Pretoria, Pretoria
Members
Palesa Rachael Lepamo, MITC Big Data Science
Supervisor(s)
Prof. Vukosi Marivate, Dr A Bosman.
Abstract
The use of solar energy is increasing globally due to the numerous benefits of solar energy. There are two types of solar energy technologies, photovoltaic and concentrating solar power. In this study, we use machine learning to detect and identify solar photovoltaic panels in aerial images of low quality. Although similar studies have been conducted, none have been applied to South African datasets. The study discusses the end-to-end framework, which includes data collection and prepossessing, a classification model to detect solar panels in images; and two segmentation models to identify the solar panels. A performance evaluation is carried out to test two state-of the-art architectures, U-net and Deeplabv3+. These architectures are usually applied to large datasets, however, in this study they are applied to a small dataset with low quality aerial images which are annotated by hand. Of the two architectures, the U-Net architecture proved to be the most suitable to our dataset.