Predicting PV Areas in Aerial Images with Deep Learning
Matthias Zech, Joseph Ranalli
Abstract
Data on the location of distributed photovoltaic installations are valuable to a variety of research activities. We have trained and applied a Fully Convolutional Neural Network to identify PV sites from aerial images of Oldenburg, Germany acquired from Google Maps. The architecture used was U-net, which was trained on a set of manually labelled images, and verified against a test dataset. The model is able to accurately estimate location and shape of PV plants in the north European town of Oldenburg. In addition, the model is able to estimate its own uncertainty, breaking the black box assumption of Deep Learning.
Topics & Concepts
Convolutional neural networkPhotovoltaic systemDeep learningArtificial intelligenceComputer scienceAerial imageBlack boxTest setSet (abstract data type)Data setArtificial neural networkArchitectureVariety (cybernetics)Machine learningData modelingPattern recognition (psychology)Image (mathematics)GeographyEngineeringDatabaseProgramming languageElectrical engineeringArchaeologySolar Radiation and PhotovoltaicsEnergy and Environment ImpactsPhotovoltaic System Optimization Techniques