Mapping of Photovoltaic Deployment in Reunion Island from Aerial and Satellite Images using the DeepSolar Framework

Abstract:

The adoption of photovoltaics by the greatest number is one of the key levers of the energy transition. Monitoring the deployment of solar installations, particularly residential ones, is an essential step to ensure the effective integration of this technology into the grid. This allows to measure the impact of public policies and to quantify the relative importance of the solar power generation source in the electricity mix. However, it is currently difficult to find complete and up to date databases containing precise locations and sizes of photovoltaic installations. Some databases, such as the Open PV Project, provide interesting insights but are often limited geographically, present low-resolution data (statistics at the level of a city or department) or are quickly obsolete. The democratization of high-resolution satellite images is the opportunity to observe the territory with unprecedented precision and frequency. Several recent works have shown the feasibility of mapping photovoltaic installations from high resolution satellite images. In 2018, Stanford University released the DeepSolar model, a Convolutional Neural Network (CNN) trained and tested on more than 300 000 solar images collected from Google Earth across the entire US.

In this work, we adapted Stanford’s DeepSolar model to produce a comprehensive map of solar installations across Reunion Island. Pleiades’s satellite image as well as high resolution airborne images were used as input for the model. Some parameters optimization and the development of additional post-processing were required to cope with regional specificities. For instance, the high occurrence of solar water heater on Reunion Island residential rooftops tricked the pre-trained model. In addition, since DeepSolar provides only the projected surface of the solar installation, as seen from the sky, we trained and tested a surface to solar capacity model based on ground truth collected on the Island.

The overall results on Pleaides images are deceiving, mainly because of the coarser resolution compared to Google Map US images, but DeepSolar exhibits a good detection performance on Reunion Island airborne images with a recall of 87% and a precision of 85%. Our surface to solar capacity model shows promising results, particularly for utility-scale installation.