Nigeria: the checkpoint mattered more than the country did¶

Nigeria was composed and run between 16 and 23 September 2026. Like Zambia, it started from nothing local: no cached imagery, no calibration quadrats, no national register, no third-party PV dataset. Unlike Zambia, it was scored against two checkpoints rather than one, and the difference between them turned out to be the most informative result of the run.

Interactive: Nigeria PV evidence atlas.

Headline¶


Verified	90.2 MWp (90% 71 to 117)
Best estimate	120.9 MWp (90% 112 to 158)
Grid	6,254 of 6,341 cells of 0.1°, 37 regions
Imagery	Sentinel-2 L2A, November 2025 to March 2026
Model	`v5_combined_france` epoch 25

A precision floor, not an estimate

Nigeria has no calibration quadrats, so there is no roofclf half and no instrument at all below the 400 m² segmentation floor. It has no glint sample, so p_unmapped is 0 and p_real collapses to the OpenStreetMap-mapped fraction, measured at 0.006 to 0.047. Recall was deliberately skipped (--recall-reference none). These are lower bounds of the kind France and Zambia publish, and they are not a statement about how much solar Nigeria has.

The finding: v4 and v5 disagree, and v5 is right¶

The chain infers twice, with v4_combined_all (the production checkpoint everywhere except France) and v5_combined_france, then scores both against the same OpenStreetMap population. Scored over 191 dissolved mapped installations at or above 400 m²:

size bin	n	v4 recall	v5 recall
100 to 500 m²	74	0.486	0.649
500 m² to 1k	54	0.648	0.889
1k to 5k	34	0.588	0.706
5k to 50k	23	0.609	0.696
above 50k	6	0.500	0.500
pooled	191	0.565	0.728

v5 is strictly dominant: it finds everything v4 finds and 31 more. v4 finds nothing v5 misses. McNemar exact two-sided p = 0.000.

This is not a small recalibration. It changes what the candidate population is:

	v4	v5
candidates	9,569	5,109
oversize, excluded from capacity	2,504	286
share of candidate area excluded	89.1%	69.1%
largest single blob	12.92 km²	4.06 km²
rooftop candidates	173	650
`est_mwp_exp`, the ceiling	1,141.8 MWp	398.0 MWp
`est_mwp_rc` rooftop	1.68 MWp	4.63 MWp
`est_mwp_rc` ground	26.36 MWp	27.22 MWp
`check-density`	32 suspect of 36	23 suspect of 37, 0 fail

The published total barely moved, from 117.5 to 120.9 MWp. Everything else did. The probability-weighted ceiling fell by a factor of 2.9 while the recall-corrected rooftop estimate rose by a factor of 2.8: v5 replaced merged false-positive sheets with actual rooftop detections. Fourteen regions now pass the ground-to-rooftop plausibility check against four before, including Lagos, Kano, Kaduna, Abuja and Rivers.

Why this matters beyond Nigeria. v5 was trained because v4 was bad over France, and its French gain was attributed to French training data. Nigeria has no training data in either checkpoint, so it is a genuine third regime, and v5 wins here too. That reframes the v5 result as at least partly a general improvement rather than a purely French one. It also means every AOI still on v4 should be re-scored before its numbers are trusted.

The v4 artifacts are kept as *_PRE_v5_20260923 so the comparison stays reproducible.

The blob problem is reduced, not solved¶

polygonize_chips merges every touching thresholded pixel with no upper bound, so a connected sheet of false positives becomes one candidate. Even under v5, 69.1% of candidate area sits above the 100,000 m² cap and is excluded from capacity by density.capacity_relevant_candidates. The largest survivor is 4.06 km².

That exclusion is what keeps the headline honest, but it is not a fix. Nigeria's failure mode is precision, not recall, and the instrument that addresses it is hard negatives mined from rejected leads, not more positives.

Plausibility gate¶

earthpv check-density reports 0 fail and 23 suspect of 37 regions. Every suspect is flagged on the ground-to-rooftop ratio with a total below 1.5 MWp, far under the check's own 50 MWp noise floor, so the ratio there is arithmetic on near-zero numbers rather than a signal. Published with the flags stated rather than suppressed.

Mapping leads¶

Nigeria is a recall-first leads product before it is a capacity product. From the v5 candidates, after removing anything within 100 m of an already-mapped OpenStreetMap solar feature, 5,074 new leads remain, of which 1,312 are at or above 5,000 m². They ship as two MapRoulette challenges in data/predictions_v5_nigeria/nigeria/:

tier	tasks	area	median	note
`leads_map`	1,026	32.2 km²	21,480 m²	5k to 100k m², ranked best first
`leads_review`	286	83.6 km²	184,814 m²	above the merge cap, redraw rather than trace

The head of the queue is confidence-1.00 rooftop leads on large roofs in Lagos, Abuja, Kano, Kaduna, Jos and Maiduguri.

Mapping from leads contaminates the next calibration

With no glint sample, mapped_frac is Nigeria's precision estimate. Mapping the model's own leads raises it mechanically, which raises p_real, which raises est_mwp_cal and est_mwp_rc. The atlas capacity would climb without the model improving at all. Freeze the pre-mapping OpenStreetMap snapshot as the evaluation reference, or filter later pulls by feature timestamp. The recall table above was computed against the 15 September 2026 snapshot, before any lead-driven mapping.

What Nigeria needs next, and what it does not¶

Not a localisation retrain. OpenStreetMap holds 1,372 Nigerian solar ways, but the median mapped feature is 225 m² and only 22% clear the 400 m² training floor. At the label-to-chip ratios observed elsewhere (Gujarat 866 features to 760 chips, Zambia 221 to 31), Nigeria yields roughly 300 to 400 chips. The v6 France experiment measured a 3,201-chip in-domain retrain landing worse than zero-shot on slope and recovery. Nigeria is an order of magnitude below the volume that already failed.

Hard negatives, mined from leads a mapper rejects, attack the 69% blob problem directly. earthpv hard-negative-chips --centers consumes them, and pakistan_hard_neg (2,626 chips) shows the approach scales.

One exhaustively mapped calibration quadrat is worth more than diffuse national mapping, because it is the only thing that unlocks a roofclf half and a real coverage ratio. A list of installations is not a calibration area: the coverage ratio needs a bounded area where the absence of PV is known. Check imagery date before drawing the box, not after mapping it. See the quadrat protocol.

Reproducing¶

pixi run python scripts/new_region.py check --iso3 NGA --name nigeria
bash scripts/nigeria_bootstrap.sh
systemd-run --user --collect --unit=earthpv-nigeria-chain \
  -p WorkingDirectory="$PWD" -p LimitNOFILE=65536:65536 -p MemoryMax=20G \
  bash scripts/run_nigeria_chain.sh

Compose reached 6,254 of 6,341 cells. The 87 that never completed are almost all in the Niger Delta and the southeast, where no Sentinel-2 scene in the November to March window clears the 30% scene cloud filter. That is a real refusal rather than a transient failure, and it is why the chain's coverage gate is set at 97% rather than 100%.

Two operational notes from the run, both now recorded in CLAUDE.md. A country-scale compose must set EARTHPV_PC_TIMEOUT_S well above the per-cell wall time when running several workers, or every cell trips the 60 s Planetary Computer patience and is downloaded twice; Nigeria livelocked for 17 hours on this and the log was indistinguishable from a provider outage. And the compose unit is memory-capped, so expect roughly one OOM restart per 10 hours, which the restart loop absorbs.