Under CopyrightMitze, FabianFabianMitze2022-12-162022-12-162022https://publica.fraunhofer.de/handle/publica/430058https://doi.org/10.24406/publica-64610.24406/publica-646In order to accompany the expansion dynamics of renewable energies in Germany and to estimate expansion potentials, precise modeling is necessary. In this work, therefore, different machine learning methods for segmenting PV plants are compared using different metrics. The methods are intended to segment PV plants as accurately as possible both in the field and on buildings. The methods to be compared are a Random Forest based model and a modified UNet. These will be trained using a large dataset covering different resolutions and plant types and then compared using an independent test dataset. With an IoU of about 84 %, the UNet shows significantly better results than the Random Forest based model with an IoU of about 30 %. The UNet was then trained again with a data set from Germany and also tested. This includes various PV installations in North Rhine-Westphalia, mainly on house roofs, but also in the open countryside. The re-trained UNet subsequently shows a IoU of about 86 % and can be used for segmentation of PV installations in Germany.deMachine LearningDeep LearningFernerkundungUNetPV-Anlagenmaster thesis