Rooftop solar is growing fast, and it plays a real role in decarbonizing the power system — it also makes communities more resilient and brings people closer to their own electricity use. But rooftop PV is a genuinely different class of asset, and a hard one to keep track of: in France, rooftop installations account for more than 99% of all grid-connected PV systems. DeepPVMapper sets out to close that gap, combining geospatial data, remote sensing, and crowdsourcing into a registry that aims to be as complete and accurate as possible. Everything it produces is released openly, kept interoperable with OpenStreetMap, and reproducible end to end by anyone who wants to check the work or adapt it. The Data and Software pages cover what the project produces and how to use it; this page is about where it came from.
DeepPVMapper grew out of a CIFRE PhD thesis — a French industrial research contract pairing an academic institution with a company — conducted between Mines Paris (PSL University) and RTE France, the French electricity transmission system operator. The thesis, “Enhancing the Reliability of Deep Learning Models to Improve the Observability of French Rooftop Photovoltaic Installations,” was defended in 2024 at Université Paris sciences et lettres (theses.fr record, summary).
The starting point was operational. In 2020, early findings — later published in RTE and the IEA’s 2021 joint report on the technical feasibility of a power system with a high share of renewables in France towards 2050 (RTE & IEA, 2021) — showed that the risk of contingencies induced by rooftop PV could increase sharply by 2050, driven by the lack of observability of these small, distributed installations. The initial goal was to characterize rooftop PV systems in a systematic way and feed that improved characterization into regional power-forecasting models.
The first step was to assemble training data and adapt an existing detection approach — building on DeepSolar (Yu et al., 2018) and 3D-PV-Locator (Mayer et al., 2022) — to the French context. That work led to the detection algorithm DeepPVMapper and the training dataset BDAPPV.
The thesis also gave an extensive treatment to the mapping algorithm’s sensitivity to distribution shift — how much its performance degrades once applied outside its training conditions. That line of work led to a new feature-attribution method, WAM, presented at ICML 2025, and to several follow-up studies on this same sensitivity in the specific case of remote-sensing-based PV detection (Environmental Data Science, 2025). The short version: PV classifiers are, more often than not, fancy grid detectors.
Coming back to the motivating question: better forecasts and reduced contingency risk for rooftop PV don’t come from a better forecasting model — they come from resolving the uncertainty around installed capacity itself. It turns out that, in France, aggregate installed capacity is tracked fairly well at the national level, but significant imbalances remain at the local scale. Reaching that conclusion required a new methodology, grounded in Bayesian statistics, to turn remote-sensing detections into a genuine estimate of installed rooftop PV capacity — see the Registry Audit.
By the time the thesis was completed, 38 French départements had been mapped — roughly 175,000 km² — making DeepPVMapper, at the time, the world’s second-largest database of its kind after the U.S.-focused DeepSolar, and the most geographically detailed.
As the rest of France was mapped, other research groups introduced their own rooftop PV detections — notably FRPV (Thebault et al., 2025). Those detections were consolidated with DeepPVMapper’s into OpenPVMapper, which aims to be a reference registry of rooftop PV installations in France — combining detectors this way measurably improves precision. The registry is updated as new imagery becomes available, and its quality keeps improving through crowdsourcing — see Contribute for how to help.
As mentioned above, rooftop PV assets are growing rapidly and present specific challenges when it comes to mapping them. DeepPVMapper and the OpenPVMapper database aim to provide a curated dataset specifically designed to address these challenges.
Rather than claiming to represent a definitive ground truth, our goal is to provide the best available information on rooftop PV, by combining remote sensing with crowdsourced data and human validation. This approach allows us to continuously improve the registry as new information becomes available.
All of it — from the training data to the final registry — remains openly accessible. Have a suggestion, or want to contribute? Get in touch.
@phdthesis{kasmi2024enhancing,
title={Enhancing the Reliability of Deep Learning Models to Improve the Observability of French Rooftop Photovoltaic Installations},
author={Kasmi, Gabriel},
year={2024},
school={Universit{\'e} Paris sciences et lettres}
}