Use of machine learning
The Deep Solaris team consists of researchers and employees from the Open University of the Netherlands, CBS, the German Federal Statistical Office Statistics Germany (DESTATIS), the Belgian Statistical Office and the State of North Rhine-Westphalia. Dr Stefano Bromuri, attached to the Faculty of Management, Science and Technology is the principal investigator on the side of the Open University of the Netherlands.
Researchers associated with the Open University investigate the extent to which models developed for other applications could also be used for the imaging of solar panels.
Dr. Deniz Iren, researcher at the Faculty of Management, Science and Technology: "We investigate several machine learning technologies for automated image classification". The technologies were assessed on a number of dimensions:
- Accuracy: How many percent of the roofs are correctly recognized (positive or negative);
- Precision: How many percent of the recognized solar panels is actually a solar panel;
- Recall: How many percent of the actual solar panels are recognized as such.
The experiments show that the models applied can be successfully used to classify solar modules from aerial photographs. The best performing models had a performance score of more than 90%.
Martine Hermans, project leader at the Faculty of Management, Science and Technology: "The result of these experiments is that two models emerged as the best. These models will be used in the continuation of the project to provide the first interactive map, also using aerial photographs of Flanders".
The project also employs research assistants that are expert in machine learning and deep learning models. In Deep Solaris Xi Chen works on creating a demo, using deep learning models, to detect solar panels in aerial images of NRW. He specializes in machine learning and deep learning as a master student of the Faculty of Data Science and Knowledge engineering (at the University of Maastricht).
Privacy measures
Statistics Netherlands always publishes data in such a way that no identifiable data about an individual person, household, company or institution can be derived from it. Although in this project public data (aerial photographs, satellite images) have been used as a basis for the analysis, the interactive maps will only show numbers of solar panels at an aggregated level.
Follow-up steps
The first interactive map is expected in spring 2019. In addition, a peer-reviewed paper is being prepared for submission in the first half of 2019. One of the next steps in the research is to investigate to what extent satellite data can be used for the detection of (large-scale) solar installations.
