This proposal concerns a thesis project at the Dept. Information Technology, UU. This thesis project will explore the possibilities of expanding a current binary segmentation of solar energy systems from aerial photos, to a multiclass segmentation. The current binary segmentation identifies either Solar Energy Systems (which includes both Photovoltaic and Solar Thermal systems) or background. The reason to expand it to a multiclass segmentation is to make the distinction between PV (Photovoltaic) and ST (Solar Thermal) systems. Making a distinction between these two is crucial for the possible application of the software by grid operators.
The dataset currently contains around 15 000 images that were taken at a height of 3 000 meters over 3 different municipalities in Sweden. The dataset will likely be expanded during the spring when new data has been annotated. In order to run simulations that actually produce some kind of meaningful result in a fast enough time, we need better computational capabilities.
The thesis is planned to end in june 2025. The computational work consists of a full dataset run of the binary classification to have as a baseline for the performance metrics decided on in the project; and then also multiple runs of the new multiclass model where different augmentations and tweakings of the program is made to optimise the performance metrics.