Namibia: a decisive localization win, and pans the model reads as panels

Namibia was added on the owner's request on 28 October 2026 as a first iteration, under the same rules as every country since the Morocco-Egypt-Yemen-South Africa queue: no calibration quadrats, no register used, the segmentation localized on Namibia's own OpenStreetMap labels and shipped only if it beats the general model on a held-out region. It does, by a wide margin, and this page runs on v17_combined_namibia.

Interactive: Namibia PV evidence atlas.

Headline

Verified	464.3 MWp (90% 336 to 657)
Best estimate	530.2 MWp (90% 480 to 664)
Grid	459 imaged cells of 0.1°, of the 7,260 that cover the country; 13 regions (geoBoundaries)
Imagery	Sentinel-2 L2A, May to September 2026
Model	v17_combined_namibia epoch 24, localized (beat v5_combined_france on the holdout; see below)

A precision floor below 400 m², and an overcount in the north

Namibia has no calibration quadrats, so there is no roofclf half and no instrument below the 400 m² segmentation floor, where household, farm and lodge systems sit. 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). Most of what the model adds in the Cuvelai basin is seasonal pans, not panels (below): about 54 of the 66 MWp by which Best exceeds Verified. For comparison, IRENA puts Namibia at 389 MW of solar PV at the end of 2025 (IRENA's own estimate), and 1,498 MWp of panels were imported directly from China from January 2017 to August 2026 (Ember). Neither is ground truth for the map.

Localizing the detector

Labels. 7,005 OpenStreetMap solar features inside Namibia: 5,889 ground, 1,110 rooftop, 6 unclassified, and no solar-thermal collectors. Most of the ground features are panel rows: Namibian mappers traced the rows of about a hundred plants one by one inside the plant outline.

Holdout. Picked from the label map by scripts/pick_val_region.py: Erongo, 20 cells, with 38 of the country's 234 dissolved positives of 400 m² and up (16.2%). Khomas, with 118, stays in training. 36 val chips, a thin holdout like Morocco's.

Corpus and recipe. 120 train chips, weighted x8 (round(7000 / 120), capped at 8), giving 960 of 27,872 train rows (3.4%). South Africa's chips join at weight 1 (3,546 train rows), as the nearest landscape with a large label set: arid Northern Cape parks and Karoo towns that no other corpus row resembles. The rest is v5's corpus, and the recipe is v5's verbatim. Training queued behind the pooled v18 run for the GPU; its first attempt was killed by systemd-oomd at epoch 6, when other jobs filled memory; the restart, from scratch, reached its best at epoch 24 (val/mIoU 0.804) and early-stopped after epoch 32, 4 hours 6 minutes on the GTX 1060.

The ship decision, scripts/compare_local_vs_v5.py on the 36 held-out chips. An installation is counted once per chip it appears in, and most are panel rows inside a few Erongo plants:

	v5	v17 (localized)
recall, installations at or above 500 m²	0.766 (723 of 944)	0.952 (899)
of which 1,000 m² and up	0.179 (20 of 112)	0.795 (89)
of which 500 to 1,000 m²	0.845 (703 of 832)	0.974 (810)
250 to 500 m² (reported, not decisive)	0.035 (12 of 346)	0.618 (214)
pixel IoU	0.456	0.662
pixel F1	0.626	0.797

Recall decides first, and the difference, 0.186, is far above 0.02: v17 ships, ahead in every size band and on outline quality. Before v17 was trained, a preview of the atlas on the released v14 was built and not published (Best 518.7 MWp); it showed the same pans.

The ground-perimeter screen

All 31 mapped ground perimeters of 50,000 m² and up (9.05 km², the largest 2.30 km²) are at least 98% imaged, and the median one is 97% detected. Three are mostly missed: a 0.23 km² outline with nothing detected, and two of 0.13 km² at 1% and 27%. No perimeter exceeds the 5 km² cap, so the screen changes no total.

Plausibility gate

earthpv check-density fails no region and marks 10 of 13 suspect on the ground-mount to rooftop ratio of the model's estimate: 4 MWp of rooftop against 136 MWp of ground, with every region below the 50 MWp floor that turns a suspect ratio into a failure. That is the missing rooftop denominator seen in every segmentation-only country: Namibia's household arrays are far below what the sensor resolves. Khomas, Omaheke and Kunene pass.

The Cuvelai pans

Best exceeds Verified by 65.9 MWp. By region, 53.7 MWp of that sits in the four regions of the Cuvelai basin: Omusati 29.2, Ohangwena 18.0, Oshikoto 5.3 and Oshana 1.2. 553 of the 965 candidates lie there, 492 of them with no building nearby. Ten of them were checked on Esri World Imagery: the six largest below the merge cap and four further down the ranking. All ten are oshanas, pans and ponds, flooded or dried, or a bare clearing; none is a solar array. The flat, bright floors and standing water of the oshanas look like ground-mounted panels to the model.

They reach Best through the precision table. With no glint sample, p_real is the mapped fraction of each size bin, and in the largest ground bin (above 50,000 m²) Namibia's mapped plants make 21 of 101 candidates real, so every unmapped outline of that size is priced at about a fifth of its area. The page states the problem in its caveats; no filter removes the pans yet. Verified, which counts only mapped installations, is not affected.

Candidate population

965 candidates: 111 rooftop, 117 ground-adjacent, 737 with no building, median 3,803 m². 69 carry a mapped OpenStreetMap geometry, 20 of them above the 100,000 m² merge cap and exempt from it. 63 model-drawn outlines exceed the cap and are excluded from capacity (15.0 km²); 52 of them (12.3 km², up to 1.2 km² each) lie in Omusati and Ohangwena, the same oshana country. The disk review page (scripts/oversize_review.py) holds the 28 that pass the lead screens. The owner judged all 28 on 3 November and found none solar, so nothing above the cap counts and no figure on this page moved.

The model's own estimate is 143.1 MWp (est_mwp_rc, 90% 87 to 231), 137.5 of it ground; thresholded and unweighted, its own outlines below the cap are 35.5 MWp (est_mwp_det).

Verified is the hand-mapped layer at the two constants. 1,254 reference features after dissolving (172 ground, 1,076 rooftop, 6 unclassified): 422.4 MWp of ground and 42.0 of rooftop. Best is floored at Verified in 34 cells.

Region	Verified (MWp)	Best (MWp)
Karas	159.9	160.3
Erongo	111.6	111.9
Hardap	65.1	66.1
Otjozondjupa	53.5	53.7
Omusati	3.3	32.5
Khomas	27.0	28.4
Oshikoto	22.1	27.4
Omaheke	17.9	18.0
Ohangwena	0.0	18.0
Caprivi	1.2	4.4
Kavango	0.0	3.2
Oshana	0.0	1.2
Kunene	0.3	0.5

Mapping leads and a calibration area

848 mapping leads, the model's outlines more than 100 m from any mapped solar feature, are in the release pv-new-leads-2026-10-28 (namibia_pv_new_leads.geojson), and the atlas links them. In the north most of them are the pans above: check each against imagery before mapping anything.

One calibration candidate passes the imagery gate, a 2 km box at Rosh Pinah in Karas (karas_s2_calib_2km), the atlas's starred area. The gate needs Esri imagery at zoom 18, the level panels are traced at, captured no more than 12 months before the imagery window. Across most of Namibia Esri serves its sharp imagery (0.3 to 0.5 m, some from January 2026) only to zoom 17, so 786 of 842 draws, and both purposive start-here boxes, were screened out. geoBoundaries' Namibian districts (ADM2) carry the wrong names (Rosh Pinah falls in "Omuthiyagwipundi", Oshakati in "Otjinene"); the regions (ADM1), which the atlas draws, are correct, and the candidate sampler names a Namibian box after its region alone.

Operational notes

The queue pulled the labels on 28 October and composed 459 cells (409 with at least 1,000 buildings, plus every cell holding a mapped solar feature) overnight, done at 05:48 on 29 October. Training then waited about nine hours for the GPU. Inference over the 459 cells took 25 minutes, and everything from postprocessing to the atlas 9 minutes.

What Namibia needs next

A filter for the Cuvelai pans. Hard negatives cut on OpenStreetMap water and wetland in the north, as Egypt's are cut on mapped pools and wastewater basins, or a seasonal-water test, before the model's figure there can be read as solar.

Map the parks and rows the model already finds, from the lead file, and the Rosh Pinah box, panel by panel. Six to ten complete areas start a small-installation half, which needs areas with zoom-18 imagery.

Reproducing

pixi run python scripts/overpass_labels_chunked.py --bbox=11.7,-29.0,25.3,-16.9 \
  --iso3 NAM --name namibia --tiles 4x4 --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 Namibia, picks the holdout and hands the country to scripts/run_country_compute.sh namibia (chips, training, the v5 comparison, inference and the atlas chain, behind marker files).