Abstract
South Africa’s rapid adoption of solar energy has led to unmonitored expansion of solar installations raising technical, financial, and safety concerns for municipalities and grid operators. While the development of methodology to monitor solar adoption flourishes worldwide, it remains underexplored in the South African context. The lack of data around solar installation makes it difficult to analyse and predict solar distribution contribution to the development of unregistered installations. This study employs a deep learning methodology using aerial imagery to assess the extent of solar distribution in South Africa, with Pretoria serving as a case study. A convolutional neural network based on InceptionV3 was trained and evaluated in four neighbourhoods of the city. An intersection-over-union score of 70.6% was achieved. The analysis revealed a faster adoption in industrial, business, and educational hubs among the four areas of interest. Lastly, city-wide inference analysis revealed a more prominent concentration in the north and south-eastern parts of the city, with a more dispersed distribution closer to the city centre and in the south-west. These findings demonstrate the feasibility of using deep learning for solar panel detection and highlight their potential in supporting sustainable energy policies and transitions in South Africa.