Yemen: solar irrigation the old labels hid, and parks that are not on the map¶
Yemen was composed and run on 11 and 12 October 2026, the second country of the
Morocco-Egypt-Yemen-South Africa queue to reach a page. Like Egypt it started
with no calibration quadrats, no national register and no third-party PV dataset to check
against, and it was run under the owner's standing instruction to localize the
segmentation on the country's own OpenStreetMap labels and to ship whichever checkpoint
scores better on a held-out region. Neither checkpoint scored anything on the holdout, and
the tie went to the localized v13_combined_yemen under that rule. Since 2026-11-02 the page
runs on v25_combined_yemen, retrained on the arrays the owner mapped from v13's leads
(below).
Read it as a map of what OpenStreetMap holds, not as an estimate of Yemen's PV.
Interactive: Yemen PV evidence atlas.
Headline¶
| Verified | 277.5 MWp (90% 238 to 325); 151.3 until 2026-11-02, 130.9 until 2026-10-30, 65.9 before the 2026-10-24 OpenStreetMap refresh |
| Best estimate | 586.2 MWp (90% 480 to 774); 457.3 until 2026-11-04, 324.7 until 2026-11-02, 302.1 until 2026-10-30, 67.1 on 2026-10-12, 99.8 on the morning of 2026-10-24 |
| of which farm ground PV, < 400 m² | 171.0 MWp (90% 85 to 335), 34 cells, calibrated on seven farm areas; 169.1 until 2026-11-02; below |
| OpenStreetMap | pulled 2026-11-01, 4,135 installations in the grid (3,315 on the late 2026-10-24 pull, 2,723 in the page before, 1,892 on 2026-10-10) |
| Grid | 810 cells of 0.1°, 22 governorates (geoBoundaries) |
| Imagery | Sentinel-2 L2A, October 2025 to March 2026 |
| Model | v25_combined_yemen epoch 30, retrained on the 2026-11-01 pull; v13_combined_yemen epoch 28 until 2026-11-02 (tied with v5_combined_france at zero on the holdout; see below) |
Household rooftops below 400 m² are not estimated
Yemen has no roofclf half: it was fitted on Yemen's nine building boxes and
fourteen farm areas on 2026-10-25 and left off, because only four boxes hold rooftop PV
and its leads were mostly greenhouses and sheds
(below), so nothing
prices rooftop PV below the 400 m² segmentation floor. Below that floor only the farm ground-PV detector counts, and only in
the 34 cells whose cropland, buildings and elevation resemble its farm areas
(below). The segmentation model has no glint sample,
so p_unmapped is 0 and p_real collapses to the OpenStreetMap-mapped fraction: 0.13 to
0.27 in v25's rooftop bins from 100 to 5,000 m² and 0.01 to 0.07 on the ground, under
0.02 everywhere for v13 (next sections). Recall was deliberately skipped
(--recall-reference none). For scale, IRENA counts 456 MW of operational PV at the
end of 2025, and the 120 MW Aden plant, in operation since July 2024, is being
doubled to 240 MW
(pv magazine, 8 September 2026).
That plant is not in OpenStreetMap. Since 2026-11-04 Best counts the model's outline of
it, which the owner judged solar: 67 MWp, the land constant on 1.34 km².
Revised 2026-10-24: OpenStreetMap refreshed, the owner's traces counted as panels, seven farm areas
The OpenStreetMap layer was pulled again (2,746 features against 1,892), the segmentation precision table and density re-derived from it, and the farm ground-PV detector refitted on seven completely mapped farm areas. Polygons the project's owner traced around the panels (869 in Yemen) now convert at the module constant, as panels, where any other mapped ground polygon converts at the land constant, as a site: about 42 MWp of Verified's rise. Earlier figures in the sections below are marked with their date.
Rebuilt 2026-10-30: the late 2026-10-24 OpenStreetMap pull
The page now counts the OpenStreetMap pull made late on 2026-10-24, after the page was
built, with the segmentation precision table and density re-derived from it: Verified
130.9 to 151.3 MWp, Best 302.1 to 324.7. The farm ground-PV figure is unchanged, the
published seven-area fit (/home/tobi/earthpv_data/yemen_7areas_20261024); it was
deduplicated against the earlier pull, so an array mapped in between can count in both.
The fourteen-area refits of 2026-10-25 (449 to 891 MWp depending on the slope veto and
the roofclf deduplication) stay unpublished. The page also gained the size chart.
Rebuilt 2026-11-02: v25, retrained on the mapped leads, and the 2026-11-01 pull
The owner mapped 827 solar features between 2026-10-25 and 2026-10-31, most of them from
v13's mapping leads. The page now runs on v25_combined_yemen, trained on the pull that
holds them, with its own precision table and density, and counts that pull: Verified
151.3 to 277.5 MWp, Best 324.7 to 457.3. The new features are 126.6 MWp of the rise
(105.6 ground, 21.0 rooftop); the model's part of Best beyond the map grew from about 4
to about 9 MWp. The farm ground-PV figure is the same seven-area fit, deduplicated again
against the new pull and v25's detections: 169.1 to 171.0 MWp. The mapping leads are
v25's (pv-new-leads-2026-11-02: 1,036 leads and 6 very large outlines, against v13's
1,073 and 25). Sections below that quote v13 keep their dates.
Revised 2026-11-04: five parks judged solar
On 3 November the owner judged, on the disk review page, the six model outlines above
the 100,000 m² merge cap that the lead screens keep: five are solar parks nobody
has mapped, 2.6 km² (the Aden plant, the park at Mocha, the Red Sea coast park and
two smaller ones in Al Hudaydah), and now count whole as ground; one is not. Best rises
from 457.3 to 586.2 MWp (90% 480 to 774); Verified did not move. Aden now fails
check-density's single-cell check, the park being 99% of the governorate's capacity:
a genuine concentration, stated rather than suppressed.
Localizing the detector¶
Labels. 1,892 OpenStreetMap solar features inside Yemen (1,733 ground, 29 rooftop, 130 unclassified points), 1.31 km² in all, median 655 m². The largest is 0.022 km². The solar-thermal filter found no solar-thermal element in the country's bounding box. 1,713 of the features sit in the Tihama plain of Al Hudaydah between 14 and 15°N, and 1,697 of those were last edited in May to August 2023: rectangles of four corners and a median 617 m², scattered across irrigated fields, the pattern of solar-powered irrigation pumps.
Holdout. Picked from the label map before training by scripts/pick_val_region.py:
'Amran Governorate, 47 cells, holding 21 of the country's 1,465 dissolved positives of 400
m² and up (1.4%). Al Hudaydah, with 1,434, is the core and stays in training. The 21
lie near 43.95°E 16.27°N, were all last edited on 23 or 24 February 2024, and are
450 to 1,755 m², 18 of them ground.
Corpus and recipe. 233 train and 13 val chips (190 with PV), weighted x8, the cap, giving 1,877 of 27,115 corpus rows (6.9%). The recipe is v5's verbatim. Best epoch 28 on the pooled validation set (val/mIoU 0.790, over every country's held-out chips, of which Yemen's are 13), early-stopped at epoch 36, 9 hours 44 minutes on the GTX 1060.
The ship decision, scripts/compare_local_vs_v5.py on the 13 held-out chips:
| v5 | v13 (localized) | |
|---|---|---|
| recall, installations at or above 500 m² | 0.000 (0 of 25) | 0.000 (0 of 25) |
| pixel IoU | 0.000 (0 true, 5 false positive pixels) | 0.000 (0 true, 0 false positive pixels) |
Both measures tie, and the rule gives a tie to the localized model. The comparison carries no information: neither checkpoint finds a single held-out pixel.
The whole-country comparison, after the atlas (compare_checkpoints_vs_mapped.py,
both checkpoints over all 810 cells, scored against the 1,465 dissolved mapped
installations of 400 m² and up), finished on 19 October, confirms the holdout:
v5 finds 6 of them and v13 2, the two parks between 5,000 and 50,000 m² for both
(McNemar p = 0.13). v13 draws 2,304 candidates below the merge cap against v5's 3,609,
larger at the median (1,799 against 1,201 m²) and over about the same area (9.2
against 9.8 km²), and 29 above it (11.2 km²) against v5's 13 (4.3 km²),
the dark-rock blobs and the three unmapped parks described below.
Retraining on the mapped leads: v25 (2026-11-02)¶
Between 2026-10-25 and 2026-10-31 the owner mapped 827 solar features in Yemen, most of them from v13's mapping leads: 514 ground and 313 rooftop, 437 of them 400 m² and up. The pull of 2026-11-01 holds 4,158 features against 3,338 on 2026-10-24; arrays of 400 m² and up rose from 1,777 to 2,197, rooftop ones from 41 to 126. The solar-thermal filter found none.
Corpus and recipe. v13's recipe and base corpus, unchanged; only Yemen's chips are new:
518 train and 55 val (430 with PV), against v13's 233 and 13, weighted x8 (29,443 corpus
rows). Best epoch 30 (pooled val/mIoU 0.790), early-stopped after epoch 38, about six hours
on the GTX 1060 (configs/terramind_pv_v25_yemen.yaml, scripts/run_yemen_v25.sh).
The holdout cannot decide. 'Amran is unchanged, but it is now almost entirely the owner's
Huth farm area: 170 mapped arrays of 400 m² and up, 162 of them ground arrays beside
fields, median 987 m², about 156 inside that calibration area. v25 finds 6, its second
epoch 3, v13 2 and v5 2; of the holdout's 5,533 PV pixels v13 marks 42 and v25 90 at 0.3
(results/backbone_compare/v25_yemen_vs_v13/). Every model is blind to these arrays, which
is what the farm ground-PV detector is for.
Against every mapped array of 400 m² and up. Both models' national probability
rasters were scored against all 2,191 dissolved OpenStreetMap arrays of 400 m² and up in
the 2026-11-01 pull, an array counting as found when any pixel touching it reaches the
threshold (scripts/yemen_v25_national_recall.py, results/yemen_v25_national_recall/).
v25 is read at 0.1, the lowest threshold at which its pixel precision reaches v13's at 0.3,
both nationally and outside both models' training images:
| Mapped arrays of 400 m² and up | n | v13 at 0.3 | v25 at 0.1 | Only v25 / only v13 |
|---|---|---|---|---|
| All | 2,191 | 0.128 | 0.152 | 58 / 6 |
| Rooftop | 126 | 0.405 | 0.460 | 8 / 1 (p = 0.04) |
| Ground, 400-1,000 m² | 1,493 | 0.036 | 0.050 | 24 / 4 |
| Ground, 1,000-5,000 m² | 558 | 0.290 | 0.335 | 26 / 1 |
| Ground, 5,000 m² and up | 14 | 0.929 | 0.929 | 0 / 0 |
| Ground mapped by 2026-10-24 | 1,713 | 0.005 | 0.012 | 12 / 0 |
| Ground added 2026-10-25 to 31 | 352 | 0.625 | 0.719 | 38 / 5 |
| Ground outside both models' training images | 224 | 0.031 | 0.067 | 8 / 0 (p = 0.008) |
| Rooftop outside both models' training images | 16 | 0.188 | 0.188 | 0 / 0 |
v25 raises half the alarms. At 0.3 it marks 100,310 pixels against v13's 204,431 and finds more arrays (305 against 280); 7.4% of its pixels touch mapped solar against 3.2%. Outside both models' training images, where little is mapped, it marks 82,000 pixels against 184,321.
How far this goes. 87% of the rooftop arrays and 83% of the ground arrays lie inside
v25's training images (v13's: 4% and 62%), so the whole-country rows favour v25; the rows
outside both models' training images are the clean test, and they point the same way on
few arrays. The arrays added in late October were mapped from v13's leads, which favours
v13. Neither model sees the Tihama rectangles mapped by 2026-10-24. Counting per candidate
instead would mislead: postprocess.replace_with_osm_geometry gives a matched candidate one
OpenStreetMap polygon, so its neighbours lose the match.
The decision. The holdout ties and both checkpoints are localized; v25 ships on the
national comparison. Its precision table matches far more candidates to mapped arrays (260
of 2,002, against 2 of 2,333 for v13), but 239 of the 260 lie inside its own training images
(outside them 21 of 1,606 match), so the higher p_real is mostly the training labels
found again; through "combine, don't add" Best takes those arrays from the map either way.
Of its 8 outlines above the merge cap (3.2 km², against v13's 29 and 11.2 km²),
three are the Aden, Mocha and Red Sea parks below. Of the six the lead screens keep, the
owner judged five parks on 3 November (those three and two smaller ones in Al Hudaydah),
which now count. check-density reported 0 fail and 7 suspect of 22 governorates, each
below the 50 MWp floor; with the parks, 1 fail (Aden, its park) and 9 suspect.
Why neither model sees Yemen's mapped PV¶
Revised 2026-10-23: the labels' geometry, not the sensor
The explanation below says the Tihama arrays are too small for Sentinel-2. A test ground in 'Amran, where 306 arrays of the same size class were traced around their panels, shows they are not: those arrays read 8 to 21% darker than their ring in the visible and near infrared. The Tihama labels are rectangles drawn around plots, and averaged over a rectangle the panels disappear. Read as bags that hold PV somewhere, the same rectangles do carry the signal, and a pixel detector trained on them finds the arrays. See Farm ground PV. The measurements below stand; the conclusion drawn from them does not.
The held-out arrays are not the exception. Scoring v13's own national inference against every mapped feature of 400 m² and up, as the ground screen does, it lights up 2 of 1,466, including the Tihama arrays it was trained on. The cause is in the imagery, not the model:
- The arrays barely differ from what surrounds them. Across the 181 Yemeni training chips with at least three PV pixels, the mean reflectance inside the mapped outlines differs from a 30 m ring around them by a median 1.6 to 2.6% per band. The same measurement gives 8 to 18% in Egypt's chips and 6 to 25% in Vietnam's. A 617 m² pump array is about six Sentinel-2 pixels of panels and gaps on a field.
- It is not a misregistration. Shifting the masks up to 60 m in every direction finds no offset at which the outlines turn dark in the red-edge and near-infrared bands; Egypt's chips, the same test, peak at zero shift.
So the localized model trained, at eight times weight, on labels it cannot resolve. It is shipped because the rule says so, not because it learned Yemen.
What the model does find: three parks the map lacks¶
v13 draws 2,333 candidates. Only two of them touch a mapped installation, so the precision
table prices every size bin, rooftop and ground, at a p_real of 0 to 0.006, and the model
adds almost nothing to either tier (next section). The detections that matter are the
largest. 29 exceed the 100,000 m² merge cap, 11.2 km² in all, and were each
inspected against the composite:
- Three are solar parks, none in OpenStreetMap: the Aden plant (44.88°E 12.88°N, 1.31 km², its panel blocks outlined almost exactly), a park of two blocks beside the power station at Mocha (43.26°E 13.37°N, 0.64 km²) and a fenced grid on the Red Sea coast north of Al Hudaydah (42.93°E 14.99°N, 0.38 km²). Together 2.33 km², about 117 MWp at the land constant.
- The other 26 are dark rock, basalt outcrops and shadowed hillsides, most of them in Sa'dah, the largest 2.54 km².
The cap excludes all 29 from capacity, since only an outline carrying an OpenStreetMap plant
geometry is exempt. That removes the rock and the parks together. All three parks are in the
exported mapping leads (yemen_pv_new_leads.geojson, Aden and the Red Sea park about
300th of 2,331, Mocha about 670th); once mapped, they enter Verified at the land constant.
Under v25 the owner judged all three solar on 3 November, and Best counts them (above).
Plausibility gate¶
earthpv check-density reports 0 fail and 18 suspect of 22 governorates. The model's
estimate is 1.2 MWp, all of it ground, so every governorate holding any of it trips the
ground-to-rooftop ratio, and none comes near the 50 MWp of ground-mount at which the ratio
counts as a failure. The gate checks the model's estimate, and here there is almost nothing
to check.
Candidate population¶
2,333 candidates: 480 rooftop, 510 ground-adjacent, 1,343 with no building, median 1,801
m². The 29 above the merge cap are 54.8% of candidate area; none carries an
OpenStreetMap plant geometry, since nothing that large is mapped in Yemen. The rest,
thresholded and unweighted, would be 191.9 MWp (est_mwp_det). Priced by the precision
table they are 1.2 MWp (est_mwp_rc, all ground, 90% upper bound 11.3), because two of
2,333 candidates match a mapped installation. In this floor a detection the map has never
confirmed counts for nothing, whether it is a real array nobody has mapped or a bright
hillside.
Verified is the hand-mapped layer at the two constants (2026-10-12: 1,891 features, 1.30 km² of ground at 65.1 MWp and 0.8 MWp of rooftop, 61.4 MWp of it the Tihama irrigation arrays). On the first 2026-10-24 pull it was 130.9 MWp: 122.6 MWp of ground and 8.3 of rooftop. On the late pull of that day it is 151.3 MWp: 135.8 of ground and 15.5 of rooftop, and Best adds the model's 7.1 MWp and the same 169.1 of farm ground PV, floored at Verified in four cells. The figures that follow are the earlier page's.
847 of the polygons are the owner's panel traces (osm_panel_tracers in
configs/aoi.yaml), 629 of them on the ground with 326,065 m², which count at the module
constant: 58.7 MWp, against 16.3 had they been priced as sites. Best adds the model's
4.2 MWp (est_mwp_rc, up from 1.2 because a few candidates now match newly mapped arrays) and
169.1 MWp of farm ground PV below 400 m², and is floored at Verified in two cells.
Operational notes¶
Compose built 810 cells and finished at 04:06 on 12 October, with no provider hand-offs and
the 10 October and 12 October compose fixes in place (scene ranking by coverage,
band-dropout check, label cells covering a whole outline; see India). Inference
over 810 cells took 43 minutes. The ground screen crashed on its first run: Yemen has no
mapped ground perimeter of 50,000 m² or more, and the screen wrote an empty table and
then failed reading its own screen column. It now writes the column header regardless, so
density reads an empty screen; the queue's three launches had all stopped at that step,
and the chain was relaunched at 15:51 once it was fixed.
Farm ground PV (added 2026-10-24)¶
Small arrays beside fields, mostly solar irrigation pumps, are counted by the
farm ground-PV detector, a pixel model, and enter Best as
their own component, beside where rooftop PV below 400 m² will sit once roofclf is
fitted here. It was first added on 2026-10-24 calibrated on 'Amran alone (32.7 MWp in 12
cells, a stated judgement interval), and recalibrated the same day on seven farm areas:
| In Best | 169.1 MWp, arrays below 400 m²: 166.5 in 32 lowland cells at Tihama's ratio, 2.6 in two highland cells at 'Amran's |
| 90% interval | 77 to 327 MWp: a bootstrap over the six Tihama areas, and a stated x/รท2 judgement for 'Amran's single area |
| Where | cells whose cropland share, building density and elevation lie within 25% of one region's areas: Tihama 24 to 85% cropland, 49 to 617 buildings per km², 54 to 295 m; 'Amran 5.3 to 8.3%, 37 to 58, 1,516 to 2,369 m |
| Not counted | every other cell; detections inside the seven areas (counted from the map) |
Why the figure is per region. The detector finds 60% of 'Amran's arrays but 6 to 50% of each Tihama area's, so the ratio of mapped panel area to detected area is 0.17 in 'Amran and 1.19 in Tihama: each detection there stands for more panel than it shows. Priced with one ratio over one box around all seven areas, the figure was 3,532 MWp in 289 cells, mostly highland terraces in Ibb, Sana'a and Ta'izz that resemble Tihama only in cropland and buildings, and more than the 2,848 MWp of panels Yemen has imported in total since 2017. Elevation separates the two landscapes (Tihama's areas lie at 68 to 236 m, 'Amran's at 1,895 m), and each cell now takes only the ratio of the region it resembles.
The figure is Tihama's hotspot ratio applied to the lowland. The owner's Tihama region was drawn around a known concentration of solar irrigation. The 32 cells priced with its ratio hold 205 m² of detections per km², under half the calibration areas' 452, which is consistent with fewer pumps elsewhere; the ratio itself is the hotspot's. It is incremental: detections within 20 m of a feature in the OpenStreetMap layer this page counts, or of a segmentation candidate Best uses, are left to those.
Rooftop PV below 400 m²: borrowing Pakistan's classifier (2026-10-28)¶
roofclf was fitted on Yemen's 23 areas on 2026-10-25 and left off this page: a median
fold AUC of 0.83, but the rooftop leads it weighted most, reviewed by hand, were greenhouse
tunnels and long poultry sheds. Yemen's mapped rooftop arrays are not too small for the
sensor (median 78 m² in Al Wehdah and 51 m² in Sayun, against 44 m² in
Pakistan's quadrats), so Pakistan's classifier, fitted on 30 quadrats and 17,151 PV roofs,
was tested on the same Yemeni building tables
(scripts/roofclf_transfer_yemen.py):
| Training | Median AUC, the 4 building boxes with PV | Median AUC, all 16 areas with PV |
|---|---|---|
| Yemen alone, leave one area out | 0.813 | 0.833 |
| Pakistan's model, unchanged | 0.809 | 0.724 |
| Pakistan and Yemen, each half the likelihood | 0.818 | 0.800 |
| Pakistan and Yemen, every row weighted 1 | 0.815 | 0.762 |
Pakistan's model ranks Yemen's roofs as well as Yemen's own. It does worse in the farm
areas, whose positives are mostly irrigation arrays credited to the nearest farmhouse, and
pooling the two countries adds about 0.01 on roofs (roof-classifier --extra-table does
this pooling). Ranking is not what keeps rooftop PV off this page.
Where the high scores land is. Pakistan's model was run over all 808 cells, and the 24 one-kilometre windows holding the most household roofs in its national top 5% were looked at on Esri imagery, nine top-scored roofs each. Six showed panels, all towns: Ibb, Yarim, Yafi', Khamir, Al Makhadir and Sa'dah. Thirteen showed none: eleven rural hamlets of stone and mud houses scored above 0.9, Ghayl Bawazir's roofless concrete shells (its top 683 roofs), greenhouse tunnels at Sahar, and the shadows of Thula's tower houses. Sana'a's windows away from Al Wehdah, where most roofs in view carry panels, hold 45 to 78 roofs in that top 5%, against 119 to 683 in the 24. Yemen's own model scores household roofs higher in its sparsest cells than in Sana'a's (mean 0.019 to 0.023 against 0.013) and puts 67% of its score-weighted roof area in cells under 100 buildings per km². None of this is measured, because only four of the nine building boxes hold PV and none is a rural hamlet with a false-positive cluster.
The first measured level: Yafi' and Qa'tabah town (2026-10-29). The first two PV-rich
town boxes to be mapped (below) are the first Yemeni test of where
the scores sit, not only how they rank. Against roofclf's own parcel label (roof_pv_area
plus parcel_pv_area, a roof with PV on 5% of it), Pakistan's unchanged model, read from
its national Yemen scoring (data/roofclf_transfer/yemen_pk/prob):
| Roofs | Buildings | With PV | AUC | Score sum per PV roof | Precision of the top-5% flag |
|---|---|---|---|---|---|
| Yafi', below 400 m² | 932 | 29 (3.1%) | 0.79 | 4.15 | 0.10, at recall 0.69 |
| Yafi', 400 m² and up | 40 | 6 | 0.69 | 3.28 | 0.20 |
| Qa'tabah town, below 400 m² | 826 | 18 (2.2%) | 0.89 | 2.74 | 0.25, at recall 0.28 |
| Qa'tabah town, 400 m² and up | 70 | 10 | 0.81 | 1.42 | 0.28 |
| Pakistan's quadrats, below / at or above 400 m² | 1.115 / 1.270 |
It ranks the two towns' roofs as well as it ranks the other boxes, and better in Qa'tabah, but on household roofs its scores sum to 3.6 per PV roof across the two (170 over 47), about 3.2 times what they sum to in Pakistan. The boxes were picked for their high scores, and a window picked that way is partly picked for its scoring errors, which pushes the ratio up. Qa'tabah town is the least affected of the ten (the fewest high-scoring roofs, 20) and still reads 2.5 times. Everything above inferred the level from where the high scores land; these two measure it.
Calibration areas¶
Ten PV-rich town boxes (candidates, 2026-10-28; Yafi' and Qa'tabah town mapped
2026-10-29). Drawn with
calib-candidates -- --aoi yemen --pv-rich: 1 km windows of at least 800 buildings per
km² with the most household roofs Pakistan's classifier scores in its national top 5%,
every window's top-scored roofs looked at on Esri imagery and kept only where panels show
(data/labels/yemen/candidate_quadrats/pv_rich_review.csv). Purposive, and chosen on a PV
signal: they add rooftop positives to the fit, and their own adoption runs high by
construction. Sana'a: Ma'in, As Sab'in, Sanhan and Bani Hushaysh; Ibb: Adh Dhihar (Ibb city)
and Yarim; Khamir ('Amran), Sa'dah, Yafi' (Lahij) and Qa'tabah town (Ad Dali'). Esri
imagery from 2024-10 to 2025-12 over every box.
Yafi', Lahij (registered 2026-10-29, Rule-1). The first PV-rich box mapped, starred on the atlas that morning by hand rather than by the classifier: Adh Dhihar in Ibb city holds the most roofs Pakistan's model scores as likely (439, against Yafi's 196), but on Esri imagery at the same zoom over the central 600 m of all ten boxes, Yafi' showed the most distinct rooftop arrays and Adh Dhihar few among cluttered roofs. Mapped panel by panel and declared finished by the owner the same day (changesets 191180704 and 191180806), against Esri Vivid imagery of 2025-10-13, inside the composite's window. It holds 40 installations, 22 rooftop and 18 ground by their overlap with VIDA footprints, 2,519 m2 in all, median 52 m2, every one below 400 m2 and 34 below 100 m2; 35 of its 972 VIDA buildings carry PV under roofclf's parcel label. The pull was confirmed against a second Overpass query (40 both times). Not yet used by any figure on this page. With Yafi' registered the star moved on by the default order, to Adh Dhihar; Yafi' showed that order is a weak guide to PV, since one in ten of the roofs it counts carried panels.
Qa'tabah town, Ad Dali' (registered 2026-10-29, Rule-1). The second PV-rich box mapped,
the town itself, 9.7 km from the rural qa_tabah_s6 box (which holds no PV). Declared
finished by the owner the same day (changesets 191181245 and 191181493), against Esri Vivid
imagery of 2025-03-21. It holds 31 installations, 23 rooftop and 8 ground,
2,142 m2 in all, median 52 m2, every one below 400 m2 and
25 below 100 m2; 28 of its 896 VIDA buildings carry PV. Confirmed by a second
Overpass query (31). Not yet used by any figure on this page.
Ghayl Bawazir, Hadhramaut (registered 2026-10-31, Rule-1, no PV). The false-alarm
town box (calib-candidates --false-alarm), starred on the atlas from 2026-10-30. Swept by
the owner against Esri Vivid imagery of 2025-03-05, within the year before the composite
window: no PV, ground or rooftop. OSM's live API (552 elements, newest edit 2025-10-09, no
power feature) and an Overpass query against the live database (OSM base 2026-10-31
08:17 UTC, 384 elements in the box, none solar) agree, so it is registered with an empty
label file, a confirmed zero. Its 951 VIDA buildings, 933 of them below 400 m2,
make it the first town box that measures the classifier's scores on roofs without solar.
Not yet used by any figure on this page. With it registered the star would have moved
on by the default order, to Adh Dhihar; the owner removed Yemen's star instead
(start_here: none, 2026-10-31).
Al Wehdah, Sana'a (registered 2026-10-23, Rule-1). The first of Yemen's two start-here
boxes (calib-candidates --priority: 1 km in the densest quarter of the country's
buildings, where the most roof area sits in buildings of 1,000 m2 or more), so a
purposive sample that runs at a higher PV rate than the country. Mapped panel by panel and
declared complete by the project owner on 2026-10-23, against Esri World Imagery (Vivid
Mosaic Metro) captured 2025-12-29, inside the composite's October 2025 to March 2026
window. It holds 87 installations, 83 rooftop and 4 ground, 14,611 m2 in
all, median 78 m2; 78 of them (90%) are below the 400 m2 segmentation
floor and 51 below 100 m2. The pull was confirmed against a second Overpass
query (88 elements both times). No second mapping pass is recorded. It is not yet used by
any figure on this page: one quadrat fits roofclf but cannot measure its skill, which
needs leave-one-quadrat-out folds.
Seven farm areas (registered 2026-10-24, Rule-1), behind the farm ground-PV figure.
- 'Amran (Huth, 129.5 km², Esri imagery June to October 2025): the test ground, 306 arrays, every one traced around its panels by the owner.
- Tihama, six areas, 331.6 km², 1,100 installations: the owner's region near Zabid, in which every installation was checked by hand. Esri's imagery there runs from August 2021 to April 2025, so the region was cut to where it shows a capture from 1 October 2024 on (12 months before the composite window), then along the 0.1° grid: At Tuhayta (104.2 km²), Al Jarrahi west, centre, east and south (53.3, 17.9, 93.3 and 32.6), and Zabid (30.3). The other 199 km² trains the detector and calibrates nothing.
238 of Tihama's installations are the owner's panel traces; most of the rest are rectangles mapped in 2023 around whole sites, which count at the land constant's share of their area (owner decision). On Esri's April 2025 imagery, a random sample of 12 rectangles all hold panels, two to four small groups covering roughly 10 to 25% of each. Every pull was confirmed by a second Overpass query; Al Jarrahi east's first pull came back 16 features short and was retried.
What Yemen needs next¶
Map the six town boxes drawn at random, in their order. calib-candidates -- --towns
(2026-10-31) drew a building at random from every building in Yemen's density grid and kept
the 1 km box around it if it holds at least 800 buildings: As Sab'in (Sana'a city), Az
Zaydiah, Dhamar city, Bani Hushaysh, Al Marawi'ah and Bayt al Faqih, 821 to 2,471 buildings
each, under Esri imagery from 2025-04 to 2025-12. No classifier score went into the draw, so
mapped complete they say what Yemen's towns hold, with or without solar; 60 of the 74
buildings drawn lay outside towns, so about one Yemeni building in five sits in one. They
lead the atlas's mapping order.
The PV-rich town boxes come after them. Yafi' and Qa'tabah town added
63 PV roofs to the 132 a roofclf fit saw in building boxes (Al Wehdah, Sayun, Bani Al
Harith and Huth), and the eight candidates left would add several times that. The two
measured the level in towns where panels show; a box in a place the classifier gets wrong
measures the false positives at the other end. That box was drawn and starred on the atlas
(2026-10-30) and mapped with no PV on 2026-10-31 (above):
ghayl_bawazir_fa1_calib_1km, written by calib-candidates --false-alarm at the first
window the PV-rich review marked no_pv. Pakistan's model puts 683 of its 966
buildings among Yemen's top 5% of household roofs; on Esri imagery of 2025-03-05 its nine
top-scored roofs are roofless or unfinished concrete shells, none with a panel. Its share
of roofs with solar runs low by design, as the PV-rich boxes' runs high.
Map farm areas outside the two regions. calib-candidates -- --farmland drew eight 3 km
farm boxes across Yemen's cropland strata (2026-10-23). The farm figure rests on 'Amran and
one Tihama region; boxes in the highland terraces of Ibb and Ta'izz, and in the lowland away
from the Zabid hotspot, are what would let it reach the rest of Yemen's farmland, and a
second highland area would replace 'Amran's judgement interval with a measured one.
Map the three parks. They are the largest PV the model sees in the country and the one change that would move Verified materially.
A holdout the sensor can resolve. Every held-out installation here is a small array in farmland or on a hillside. A holdout that cannot be detected by either checkpoint cannot decide between them; until Yemen has a mapped park of its own, the ship rule here is a coin toss with a fixed outcome.
Reproducing¶
pixi run python scripts/overpass_labels_chunked.py --bbox=41.8,12.1,54.6,19.0 \
--iso3 YEM --name yemen --tiles 4x3 --timeout 300 --retries 8
systemd-run --user --collect --unit=earthpv-country-queue -p WorkingDirectory="$PWD" \
-p LimitNOFILE=65536:65536 -p MemoryMax=4G bash scripts/run_country_queue.sh
The queue pulls the labels, runs the solar-thermal filter, composes Yemen, picks the
holdout and hands the country to scripts/run_country_compute.sh yemen (chips, training,
the v5 comparison, inference and the atlas chain, behind marker files).
The 2026-11-02 retrain on the 2026-11-01 pull (v25), on top of that:
pixi run python scripts/overpass_labels_chunked.py --bbox=41.8,12.1,54.6,19.0 \
--iso3 YEM --name yemen --tiles 4x3 --timeout 300 --retries 8 --fresh
pixi run python scripts/drop_solar_thermal.py --aoi yemen --by-id
systemd-run --user --collect --unit=earthpv-v25-yemen -p WorkingDirectory="$PWD" \
-p LimitNOFILE=65536:65536 -p MemoryMax=14G bash scripts/run_yemen_v25.sh
pixi run python scripts/prepare_national_osm_solar.py --aoi yemen \
--labels data/labels/yemen_overpass_solar.parquet \
--keep-detected-screen results/yemen_osm_ground_screen/screen_full.csv
F=/home/tobi/earthpv_data/yemen_7areas_20261024 # the published seven-area farm fit
pixi run earthpv groundclf-capacity --aoi yemen --score-dir $F/blobs --calib-dir $F \
--out-dir data/predictions_v25/yemen/groundclf_capacity --pred-dir data/predictions_v25 \
--osm-solar data/labels/yemen_national_osm_solar.parquet
pixi run earthpv atlas --aoi yemen --pred-dir data/predictions_v25 \
--osm-solar data/labels/yemen_national_osm_solar.parquet --include-offgrid-osm \
--imagery-date-range "Oct 2025 - Mar 2026" \
--ground-cells data/predictions_v25/yemen/groundclf_capacity/ground_cells.parquet
pixi run python scripts/release_leads.py --aoi yemen --out <release directory> --upload
pixi run python scripts/yemen_v25_national_recall.py
The 2026-10-24 refresh, on top of the first run (farm areas registered with
new_calibration_quadrat.py --aoi yemen --farm --geojson <area>):
pixi run python scripts/overpass_labels_chunked.py --bbox=41.8,12.1,54.6,19.0 \
--iso3 YEM --name yemen --tiles 4x3 --timeout 300 --retries 8 # after moving the tile cache aside
pixi run python scripts/drop_solar_thermal.py --aoi yemen --by-id
# screen, calibrate-candidates, density, check-density, prepare_national_osm_solar.py: as in
# run_country_compute.sh
pixi run earthpv groundclf --aoi yemen \
--weak-region data/labels/yemen/tihama_stale_imagery_region.geojson \
--osm data/labels/yemen_overpass_solar.parquet
pixi run earthpv groundclf-score-national --aoi yemen
pixi run earthpv groundclf-capacity --aoi yemen --osm-solar data/labels/yemen_national_osm_solar.parquet
pixi run earthpv atlas --aoi yemen --osm-solar data/labels/yemen_national_osm_solar.parquet \
--include-offgrid-osm --imagery-date-range "Oct 2025 - Mar 2026" \
--ground-cells data/groundclf/yemen/capacity/ground_cells.parquet