Poland: a strong holdout, a clear localization win, and parks the map has not caught up with

Poland was added on the owner's request on 19 October 2026 and run under the same rules as the Morocco-Egypt-Yemen-South Africa queue: no calibration quadrats, no register used, the segmentation localized on Poland's own OpenStreetMap labels and shipped only if it beats the general model on a held-out region. It does, on a large holdout, and this page runs on v15_combined_poland.

Interactive: Poland PV evidence atlas.

Headline

Verified	6,604.3 MWp (90% 4,618 to 9,528)
Best estimate	9,110.3 MWp (90% 7,248 to 12,083)
Grid	3,948 cells of 0.1°, 16 voivodeships (geoBoundaries)
Imagery	Sentinel-2 L2A, May to September 2026
Model	v15_combined_poland epoch 27, localized (beat v5_combined_france on the holdout; see below)

Revised on 4 November

Until 4 November this page read Best 7,780.9 MWp (90% 6,411 to 10,398). On 3 November the owner judged, on the disk review page, every model outline above the 100,000 m² merge cap that the lead screens keep: 113 of the 147 are solar parks nobody has mapped, 26.6 km², and now count whole as ground; 29 are not, and 5 left unsure stay out until they are settled. Best rises by 1,329.4 MWp; Verified did not move.

A precision floor, and blind to most of Poland's solar

Poland has no calibration quadrats, so there is no roofclf half and no instrument 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, and recall was deliberately skipped (--recall-reference none). Poland had about 21.8 GW installed by February 2025 (Wikipedia), most of it household systems far below the floor. Panels imported directly from China, 21.8 GW from January 2017 to August 2026 (Ember), are not a ceiling here: the installed fleet already exceeded them in early 2025, because much of Poland's Chinese supply arrives through other EU countries.

Localizing the detector

Labels. 67,110 OpenStreetMap solar features inside Poland, of which 131 are solar-thermal collectors (0.13 km²), removed before chips were cut. That leaves 66,979 (36,998 ground, 21,963 rooftop, 8,018 unclassified). Every Overpass mirror failed the filter's usual country-wide scan, so it ran as 66 one-degree tiles.

Holdout. Picked from the label map by scripts/pick_val_region.py: Lublin Voivodeship, 301 cells, with 359 of the country's 4,267 dissolved positives of 400 m² and up (8.4%). Greater Poland, with 640, stays in training. Because of a grid-origin mismatch fixed after the run (below), the held-out cells are Lublin's shifted one 0.1° column east. Holdout, buffer ring and training set all use the same cell names, so the split stays clean: it is a contiguous region 7 km east of the one intended. 706 val chips, 361 of them with PV, 7.3% of Poland's labelled PV pixels.

Corpus and recipe. 7,666 train chips, weighted x1 (round(7000 / 7,666)), giving 8,515 of 33,753 corpus rows (25.2%). The recipe is v5's verbatim. The first training run was killed by systemd-oomd at epoch 8, when another session's jobs filled memory; the restart reached its best at epoch 27 (val/mIoU 0.804) and early-stopped at epoch 35, 12 hours 35 minutes on the GTX 1060.

The ship decision, scripts/compare_local_vs_v5.py on the 706 held-out chips:

	v5	v15 (localized)
recall, installations at or above 500 m²	0.927 (3,192 of 3,444)	0.960 (3,306)
of which 1,000 m² and up	0.826 (683 of 827)	0.917 (758)
of which 500 to 1,000 m²	0.959 (2,509 of 2,617)	0.974 (2,548)
250 to 500 m² (reported, not decisive)	0.836 (744 of 890)	0.954 (849)
pixel IoU	0.571	0.645
pixel F1	0.727	0.784

Recall decides first, and the difference, 0.033, is above 0.02: v15 ships, ahead in every size band and on outline quality.

A checkpoint-selection bug, caught before publication. The killed run's checkpoints stayed in the run directory, and Lightning wrote the restarted run's state to last-v1.ckpt. scripts/pick_best_checkpoint.py read only last.ckpt and picked the dead run's epoch 7 (0.776); the decision, inference and postprocessing all ran on it before this was noticed at the density step. The picker now reads every run's state (South Africa's and Morocco's picks are unchanged), and the chain was rerun from the pick. On the wrong checkpoint the holdout read recall 0.965 against v5's 0.927 but pixel IoU 0.571 against 0.571; epoch 27 trades a little recall for much better outlines.

The ground-perimeter screen

All 436 mapped ground perimeters of 50,000 m² and up (72.5 km², the largest 3.1 km²) are fully imaged, and 432 are detected, at a median 99% of their area. The four with nothing detected are 0.05 to 0.11 km². No perimeter exceeds the 5 km² cap.

The screen first read every Polish perimeter from the wrong file. compose snaps the grid to the country's own lower-left corner, and Poland's border starts at 14.12°E, so its cells are named from 14.1°E, not the configured 14.0. density replicates that; the screen and the holdout picker used the configured origin, read each perimeter from the neighbouring cell's composite, and called 398 of 436 "insufficient imagery". Both now use density._grid_origin. No other published country is affected (their snapped and configured origins agree), and no Polish total moves, since the screen only changes capacity for perimeters above 5 km².

Plausibility gate

earthpv check-density fails all 16 voivodeships on the ground-mount to rooftop ratio of the model's estimate: 5 MWp of rooftop against 4,709 MWp of ground. That is the missing rooftop denominator seen in France and India: Poland's household arrays are far below what the sensor resolves.

Candidate population

8,995 candidates after four inside solar-thermal outlines were removed: 823 rooftop, 2,616 ground-adjacent, 5,556 with no building, median 13,608 m². 155 oversize candidates carry a mapped plant outline and are exempt from the 100,000 m² merge cap. 168 model-drawn ones exceed it and are excluded from capacity (39.7 km², 18% of candidate area, the largest 1.31 km²), and in Poland most of them are real: all eight of the largest and at least seven of the eight around rank 40 are solar parks on the composite, six of the largest eight with no mapped solar within 100 m. At the land constant they would be about 2 GWp. The cap is kept, as for every country, because it is what removes the rock, fields and pits elsewhere; these parks are in the exported mapping leads (poland_pv_new_leads.geojson, 5,793 leads) and enter Verified once mapped. On 3 November the owner judged the 147 that the lead screens keep: 113 are parks (26.6 km²) and now count whole as ground, 29 are not, and 5 left unsure stay out until settled.

The model's own estimate is 6,071.3 MWp (est_mwp_rc), 4,741.9 (90% 3,124 to 6,969) before the 113 judged parks, nearly all of it ground and about half of it (2.2 GWp) the 155 mapped outlines it reproduces; thresholded and unweighted, its own outlines below the cap are 385 MWp (est_mwp_det).

Verified is the hand-mapped layer at the two constants. 44,551 reference features after dissolving (14,613 ground, 21,920 rooftop, 8,018 unclassified): 6,522.8 MWp of ground and 81.5 of rooftop. Best is floored at Verified in 1,174 cells.

Voivodeship	Verified (MWp)	Best (MWp)
Greater Poland	1,462.4	1,825.8
West Pomeranian	620.6	837.2
Lubusz	489.8	787.3
Pomeranian	551.9	774.9
Lower Silesian	464.2	736.5
Masovian	377.6	611.8
Warmian-Masurian	394.1	530.1
Lublin	474.0	526.3
Kuyavian-Pomeranian	342.4	430.6
Lodz	210.2	390.3
Podlaskie	309.2	366.4
Subcarpathian	274.3	329.3
Swietokrzyskie	174.8	257.9
Opole	162.9	256.4
Lesser Poland	111.3	191.7
Silesian	124.6	180.2

Operational notes

Compose built 3,948 cells (3,793 by building density, 2,847 holding a mapped solar feature, the two overlapping) from 19 October 09:57 to 21 October 06:33, at 60 to 110 cells an hour. The holdout pick then failed every queue tick for 15 hours, because pick_val_region.py --write only fills an empty val_tiles: [] and new_region.py add never wrote one; it does now. Inference over 3,948 cells took 2 hours 13 minutes, density 43 minutes.

What Poland needs next

Map the parks the model already outlines. scripts/park_leads.py --aoi poland writes them as a verification package: 351 model detections of 5 ha and up with no mapped solar within 100 m, about 2.4 GWp at the land constant, largest first, with iD and JOSM links, suggested tags, a MapRoulette file and a contact sheet. Of the 24 largest, 23 are solar parks on the composite and one is a quarry. Each needs a person to confirm it against high-resolution imagery and trace the site; the model outlines only mark the place.

A rooftop instrument. Poland's household systems are the bulk of its fleet and all below the floor; an area mapped panel by panel would be the start of a small-installation half.

Reproducing

pixi run python scripts/overpass_labels_chunked.py --bbox=14.0,49.0,24.2,54.9 \
  --iso3 POL --name poland --tiles 5x4 --timeout 300 --retries 8
systemd-run --user --collect --unit=earthpv-country-queue -p WorkingDirectory="$PWD" \
  -p LimitNOFILE=65536:65536 -p MemoryMax=4G bash scripts/run_country_queue.sh

The queue pulls the labels, runs the solar-thermal filter, composes Poland, picks the holdout and hands the country to scripts/run_country_compute.sh poland (chips, training, the v5 comparison, inference and the atlas chain, behind marker files).