Nationally Consistent, Locally Incomplete: A Bayesian Remote-Sensing Audit of Rooftop Photovoltaic Registries
Gabriel Kasmi, Yves-Marie Saint-Drenan, Laurent Dubus, Philippe Blanc
stat.AP · cs.LG
Quick verdict
Bayesian-corrected satellite audit matches France's PV registry nationally (3.3%) but finds local gaps up to 61%; rigorous; recall ground truth is OSM-based.
Summary
The paper reverses the usual validation arrow: instead of benchmarking satellite-based rooftop PV detections against official registries, it turns an imperfect detector into a calibrated instrument and audits the registry. Applied to continental France for the ≤36 kWp segment, the pipeline (DeepPVMapper on 20-cm IGN orthoimagery; 589,623 detections totalling 2.90 GWp raw) is corrected via Bayesian posteriors over precision (0.82) and pipeline-level recall (0.61), estimated from 31,870 manual annotations, yielding 4.03 GWp [3.96–4.11] (99% CI) against a date-matched TSO registry of 3.90 GWp — a 3.3% national agreement. Locally, 25 departments are flagged under-reported (228.2 MWp) and 8 over-reported; 18 units (172 MWp) survive a specification battery ('hard core'), with registry coverage as low as 39% in Corse-du-Sud. A municipal fixed-effects regression attributes 45% of hard-core missing capacity (77/172 MWp) to reporting through local distribution companies, which show 42% lower registry coverage. A privacy-driven truncation bias in the public registry (14,019 municipalities censored) is quantified. Key limitations: recall ground truth is seeded by OSM with untestable rural selection; the surface-to-capacity coefficient is the largest sensitivity; detection shows physical presence, not grid connection; two results are not reproducible due to disclosure constraints.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Bayesian-corrected satellite detections estimate 4.03 GWp [3.96–4.11] (99% CI) of ≤36 kWp rooftop PV in continental France, agreeing with the date-matched TSO registry (3.90 GWp) within 3.3% nationally.
- Registry coverage falls as low as 39% of the estimated fleet (Corse-du-Sud); 25 units are under-reported (228.2 MWp) and an 18-unit, 172 MWp 'hard core' survives every specification tested.
- Reporting through local distribution companies (ELDs) explains 45% (77/172 MWp) of hard-core missing capacity — presented as the first empirical evidence that crossing operator reporting interfaces degrades registry completeness.
- The public registry (RNI) carries a privacy-driven truncation bias censoring 14,019 municipalities (40.3%), flagging 32 units on the truncated view versus 8 after correction.
- The audit protocol is exportable: annotation effort is pre-computable via a closed-form budgeting rule, and dating-granularity sensitivity equals accumulated local growth.
Strengths
Validation arrow reversed with a national sanity check
Corrected detections (4.03 GWp, 99% CI 3.96–4.11) and the date-matched TSO registry (3.90 GWp), built from entirely disjoint data, converge within 3.3% — the first national-scale test where remote sensing audits the registry rather than the reverse.
Pipeline-level error measured, not borrowed from test sets
31,870 manual annotations (18,023 precision, floor 120/unit; 13,847 recall, floor 80/unit) expose the 0.84 model-level vs 0.61 pipeline-level recall gap; the delta-method budgeting rule (Fig S2) lets future audits price annotation effort in advance.
Conservative, multi-layer robustness architecture
Credible identification of the fragmentation mechanism
Unusual transparency on reproducibility and conflicts
Weaknesses
Recall ground truth seeded by OSM with untestable selection
Detection certifies physical presence, not grid connection
National convergence is coefficient-fragile
Two key results not independently reproducible
Conflict of interest with the audited institution
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper flips the standard evaluation direction in distributed-PV mapping: instead of validating satellite detections against an official registry, it converts an imperfect detector into a Bayesian measurement instrument and audits the registry with it. Five claims are advanced: (1) corrected detections give 4.03 GWp [3.96–4.11], 99% CI, for the ≤36 kWp segment, within 3.3% of the date-matched TSO registry (3.90 GWp); (2) locally, 25 of 93 reporting units are under-reported (228.2 MWp; worst coverage 39% in Corse-du-Sud) and 8 over-reported (99.4 MWp), with an 18-unit, 172 MWp 'hard core' surviving every specification; (3) reporting via local distribution companies (ELDs) explains 45% (77/172 MWp) of hard-core missing capacity; (4) the public RNI carries a quantified privacy-driven truncation (14,019 municipalities censored); (5) the protocol exports elsewhere with a pre-computable annotation budget.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How much rooftop solar does France actually have, according to this study?
For installations at or below 36 kWp, the corrected satellite-based estimate is 4.03 GWp (99% credible interval 3.96–4.11 GWp), within 3.3% of the grid operator's (RTE) connection registry, which stands at 3.90 GWp after aligning connection dates to the aerial imagery dates.
What problems did the study find in French PV statistics?
Three things. Locally, 25 of 93 departments report less small-rooftop-PV capacity than the corrected estimate implies (coverage as low as 39% in Corse-du-Sud). Reporting through small local distribution companies shows 42% lower registry coverage and explains 45% of the robust missing capacity. And the public open-data registry hides capacity in 14,019 small municipalities due to a privacy rule that removes municipalities with fewer than ten installations.
Can this audit method be used in other countries?
In principle yes: it needs high-resolution orthoimagery, any rooftop-PV detection model, and a bounded manual annotation effort whose size can be computed in advance from a formula. The authors propose the United States (fragmented reporting across 72 organisations) and Germany (direct owner registration) as natural test cases, and Pakistan — where registry figures (5–6 GW) and independent estimates (27.5–33.5 GW) diverge several-fold — as a case where corrected remote sensing would be the only measurement.
Who conducted the study and is there a conflict of interest?
Researchers at MINES Paris (PSL) and RTE, the French transmission system operator; two authors are or were RTE-affiliated, and RTE's registry is one of the datasets audited. The design mitigates this — corrections are direction-agnostic, an over-reporting negative control is maintained, and the findings include under-reporting in RTE's own data — but the raw RTE extract cannot be redistributed, so two results cannot be independently reproduced from the released data.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Several passages were omitted in text extraction (parts of Results 'Lessons', Methods 'Practical implementation', the specification-battery narrative, and part of Note S2); judgments on those areas rely on cross-references elsewhere in the text.
- Figures are available only as extracted captions and axis labels; map panels (Figure 1, S3, S9) and scatter plots could not be visually inspected.
- Supplemental derivations (e.g., the delta-method budgeting rule behind Figure S2) are partially truncated and were not independently re-derived.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.