Identifying and counting commercial solar installations in South Africa
Ryan Shackleton , Matthew Lazenby
Abstract: The objective of this project is to successfully detect, map and determine the solar panel density within the Tshwane Municipality region. The project was carried out in conjunction with the University of Pretoria's Department of Geography, Geoinformatics and Meteorology. The department provided external satellite images of the Tshwane Municipality area. To achieve the stated objectives, three different machine learning models were developed based on the pixel-wise classification methodology. The three models namely Support Vector Machine, Random Forest and Neural Network achieved accuracies of 91.25%, 94.34% and 93.37% respectively. All three models successfully detected and mapped the presence of solar panels, as well as determined the solar density, when provided a satellite image. A local web application was developed to showcase the results of the research. The Random Forest model resulted in the highest accuracy, quickest training and prediction time and therefore, is the most optimal. The biggest limiting factor within the project is the extensive computing power required.