2020-09-21, 15:50–16:20, Room 2
A novel method for detecting solar panels and its geometry on aerial imagery is presented. Deep Learning with PyTorch is being used for segmentation. The goal is to know the exact locations, dimensions and potential of every solar installation in Switzerland.
Sol AI is a prototypical framework solution to provide location and geometry for millions of solar panel installations in Switzerland. We programmed a big data approach using parallelized multi gpu deep learning with Pytorch and additional post-classifiers on large scale 10cm SWISSIMAGE aerial imagery datasets by swisstopo. It is shown how we found a solution to extract geometries and energy potential and how to distinguish different solar panel types. Data handling, labelling strategies, geoconversions, read/write NoSQL interfaces, as well as deep learning training and inferencing results are described. Furthermore we show challenges, sample data, optimization strategies and focus on the computational demands the large amount of data imposes.