Overview of the DeepPVMapper approach for detecting and characterizing rooftop PV systems
A Remote Sensing-Based Method for Mapping Rooftop Photovoltaic Systems
As the deployment of small-scale rooftop photovoltaic (PV) systems accelerates, transmission system operators (TSOs) face growing challenges in accessing reliable and comprehensive data on installed PV capacity. In France, 99% of PV grid connections correspond to rooftop systems connected to distribution networks. Yet, the accuracy and uncertainty of national PV registries remain undocumented, limiting their use in critical grid operations. In this context of sustained growth, timely and verifiable data are essential to ensure forecasting accuracy and secure grid management. We introduce DeepPVMapper, a deep learning pipeline trained on aerial orthoimagery to detect rooftop PV systems and estimate their installed capacity, tilt and azimuth angles. Applied at the French national scale, the method identifies over 500,000 installations and generates a geospatial dataset to assess and audit the completeness of official registries. This study focuses on empirically evaluating the added value of Earth Observation and deep learning methods in practice. Using spatial aggregation and inter-source consensus metrics, we analyze how DeepPVMapper outputs align or diverge from the French TSO registry. Results reveal strong correlations in periurban areas, suggesting that existing registries are generally well-structured at aggregate scales. However, we also identify persistent gaps in rural zones where EO-based mapping highlights systematic omissions or errors. This work demonstrates how remote sensing acts as an independent, scalable lens to benchmark and improve the quality of PV datasets. It contributes to assessing the operational relevance of EO-based tools for energy system monitoring where direct observability is limited.
The algorithm processes high-resolution aerial imagery through a two-stage pipeline:
The extracted polygons are converted to geolocated GeoJSON files, and system characteristics (tilt, orientation, installed power, surface area) are computed using auxiliary data sources. The system applies various filters to ensure quality, including restrictions to installations under 36 kWc located on rooftops.
The DeepPVMapper system has been validated against the French national registry of installations (RNI), demonstrating reliable performance in mapping distributed rooftop photovoltaic systems. The algorithm successfully processes large-scale aerial imagery datasets and generates comprehensive inventories of PV installations with detailed characteristics.
@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}
}[1] Kasmi, G. (2024). Enhancing the Reliability of Deep Learning Models to Improve the Observability of French Rooftop Photovoltaic Installations. PhD Thesis, Université Paris sciences et lettres.
[2] Kasmi, G., Dubus, L., Blanc, P., & Saint-Drenan, Y. M. (2022). Towards unsupervised assessment with open-source data of the accuracy of deep learning-based distributed PV mapping. arXiv preprint arXiv:2207.07466.