# 07 — Land use: is the KDMP standing on a graveyard, or in a paddy field? **Run**: `python reports/07-landuse-polygons/run.py` · Samples cloud rasters over HTTP (`--skip-rasters` to go fully offline) · **Last run**: 2026-08-13 (on 08-13 coordinates) **Source**: `data/osm/indonesia-latest.osm.pbf` (Geofabrik, 2026-08-07) + `kopdes_locations.csv` > **Candidates, not findings** — the same rule as [04](../04-siting-screen/). > Read [What this cannot settle](#what-this-cannot-settle) before citing any row. ## Why this screen exists Two accusations in the public critique are invisible to every analysis we have run so far: - **_"dibangun di tanah kuburan"_** — built on burial ground. ESA WorldCover has no cemetery class at all, so [04](../04-siting-screen/) is structurally blind to it. OSM has the polygons. - **_"dibangun di tengah sawah"_** — built in the middle of a paddy field. 04 found **exactly one** cropland case in 2,500 candidates, and said so as a known limit: it ranks by _isolation_, and a shop in a paddy field sits right next to a village by construction. This screen has no isolation term, so it can see them. Both verdicts came back — and they point in opposite directions. ## Method — point-in-polygon, and two ways to be wrong about distance Different machinery from 04. These are vector polygons, not raster pixels. 1. **One pass over the 1.73 GB PBF**, assembling closed ways and multipolygon relations into areas, with the tag filter applied on the C++ side in both the area-candidate pass and the output stream. **258,224 features in 100 s** → `data/osm/indonesia_landuse.parquet` (50 MB, cached; rebuild with `--rebuild-index`). Iterating tags in a Python callback instead is the 92-minute mistake recorded in `AGENTS.md`. 2. **An STRtree per class.** `query(predicate="within")` answers "is this KDMP inside a cemetery?" for all 83,379 points at once. 3. **Distance in two steps.** The tree ranks candidates in degree space — near the equator that is within ~2% of isotropic, fine for _which_ polygon is closest and useless as a published number. The winner is then re-measured **geodesically** (`pyproj.Geod`) against the actual nearest point on the geometry. Nothing is ever buffered in degrees: 0.01° is a different distance in Aceh than in Papua. | Class | OSM features | Polygons | Points | | ----------------------------------------- | ------------ | -------- | ------ | | `landuse=farmland` | 77,075 | 77,038 | 37 | | `amenity=place_of_worship` | 89,339 | 66,023 | 23,316 | | `place=village` | 76,973 | 1,128 | 75,845 | | `landuse=cemetery` + `amenity=grave_yard` | 9,095 | 8,413 | 682 | | `amenity=marketplace` | 5,742 | 4,270 | 1,472 | ## The comparator, without which none of this means anything "2,209 cooperatives are inside farmland" is not a finding on its own. If village institutions in general land inside mapped farmland at that rate, the number describes how OSM draws Indonesia, not where KDMP were built. **OSM's own `place=village` nodes** are the natural comparator: they are what "the right spot for a village institution" looks like, placed by mappers with no stake in this argument, at roughly the same national density (75,845 vs 83,379). The identical screen runs over them. [`null_comparison.csv`](null_comparison.csv) | Inside a… | KDMP | OSM village nodes | Excess | | ------------------- | ------------------ | ----------------- | -------------------- | | farmland polygon | **2.649%** (2,209) | 1.097% (832) | **+1.55 pts — 2.4×** | | cemetery polygon | 0.026% (22) | 0.022% (17) | +0.004 pts — **nil** | | marketplace polygon | 0.019% (16) | 0.021% (16) | −0.002 pts — nil | A second comparator — **10,580 mapped minimarkets**, sited by companies that care only about footfall — is in the same file and looks far more dramatic (0.151% against a province-reweighted KDMP rate of 3.399%, a 22× gap). **Do not quote it.** Province-reweighting fixes the between-province confound and does nothing about the one that matters: within any province, minimarkets are in towns and KDMP are in desa. It belongs in the file as an upper bound, not in a sentence. ## Finding 1 — the paddy-field claim survives, with a nameable shortlist [`farmland_funnel.csv`](farmland_funnel.csv) · [`farmland_candidates.csv`](farmland_candidates.csv) Being inside a farmland polygon is not enough on its own, so the funnel strips out everything that could be an artefact: | Step | Cooperatives | | ----------------------------------------------------------- | ------------ | | inside a mapped farmland polygon | 2,209 | | ≥ 100 m from the field edge (in it, not on the verge of it) | 1,027 | | … and ≥ 100 people within 1.4 km (the ones 04 discards) | 1,027 | | … and the polygon is not a coarsely-drawn whole plain | **538** | That last step removed **489 of 1,027** — nearly half. Some contributors trace an entire agricultural landscape as a single `landuse=farmland` way, settlements included, and a point 3 km "deep" inside one of those is in a village, not a field. Any polygon containing 2 or more OSM village nodes is dropped for that reason. Without this check the headline number would have been twice as large and wrong. **Independent confirmation.** The OSM polygon is one person's tracing. ESA WorldCover is a 10 m satellite classification produced with no knowledge of OSM. Sampling all 538 candidate points (reusing 04's COG range-request sampler): | Land cover at the point | Candidates | | ---------------------------------- | --------------- | | **Cropland** | **448 (83.6%)** | | Tree cover | 50 | | Grassland | 25 | | Built-up | 7 | | Water / wetland / shrub / mangrove | 6 | **448 cooperatives are placed on agricultural land by two independent sources, at least 100 m from the nearest edge of the field, in a desa with people living around them.** 209 of them carry a land-asset status of `Terverifikasi` — the land was officially signed off. The geography is not where 04 looked at all: **Aceh (92), Jawa Tengah (60), Jawa Timur (58)** lead the confirmed-and-roadside set, against 04's Papua-and-Maluku concentration. These are two different phenomena, and only this screen sees the second one. ### The ten deepest, confirmed by both sources, next to a mapped road | Cooperative | District | Depth into field | Polygon | Land status | | -------------------------- | ------------------------------ | ---------------- | ------- | ------------------- | | KDMP KARANGSEGAR PEBAYURAN | Jawa Barat, Kab. Bekasi | **1,067 m** | 28 km² | Sedang Diverifikasi | | KDMP CERUKCUK | Banten, Kab. Serang | 1,027 m | 28 km² | Sedang Diverifikasi | | KDMP BENA | NTT, Kab. Timor Tengah Selatan | 974 m | 14 km² | Sedang Diverifikasi | | KDMP SUKAMURNI SUKAKARYA | Jawa Barat, Kab. Bekasi | 735 m | 17 km² | **Terverifikasi** | | KDMP PILANGSARI JATITUJUH | Jawa Barat, Kab. Majalengka | 708 m | 117 km² | Sedang Diverifikasi | | KDMP SUKAWANGI | Jawa Barat, Kab. Bekasi | 693 m | 17 km² | Tidak Ada Lahan | | KDMP PASURUHAN | Jawa Tengah, Kab. Pati | 691 m | 14 km² | — | | KDMP SAMBIREJO SARADAN | Jawa Timur, Kab. Madiun | 620 m | 4.6 km² | Dipertimbangkan | | KDMP SRIKATON KAYEN | Jawa Tengah, Kab. Pati | 608 m | 14 km² | **Terverifikasi** | | KDMP GROBOGAN | Jawa Timur, Kab. Madiun | 576 m | 2.3 km² | **Terverifikasi** | Every row in [`farmland_candidates.csv`](farmland_candidates.csv) carries an `imagery_url`. ## Finding 2 — the burial-ground claim does not hold up [`cemetery_candidates.csv`](cemetery_candidates.csv) **22 cooperatives out of 83,379 sit inside a mapped burial ground — a rate statistically indistinguishable from where OSM puts its own village centres** (0.026% against 0.022%; 22 cases against 17). And the 22 do not survive inspection as a group: - **Only 7 are more than 25 m inside**, and only 2 more than 100 m. The rest are boundary effects — a kelurahan office next to the wall of a cemetery. - **16 of the 22 are in a _kota_, not a _kabupaten_.** These are large urban TPU polygons in Jakarta, Surabaya, Semarang, Bengkulu — dense places where a public cemetery is a city block and everything is next to everything. | Cooperative | Burial ground | Depth | Land status | | -------------------------------- | --------------------------------------- | --------- | ------------------- | | KDMP MARGAMULYA TELUKJAMBE BARAT | **San Diego Hills Cemetery** (Karawang) | **481 m** | **Terverifikasi** | | KOP. KEL. SEMPER TIMUR | TPU Semper (Jakarta Utara) | 129 m | Sedang Diverifikasi | | KOP. KEL. PANDAN KASTURI | Ambon War Cemetery | 82 m | Sedang Diverifikasi | | KOP. KEL. SEMABUNG BARU | Pekuburan Sentosa (Pangkal Pinang) | 82 m | Sedang Diverifikasi | | KDMP BARON | Makam Bong Cina Baron (Nganjuk) | 80 m | **Terverifikasi** | | KOP. KEL. PUTAT JAYA | TPU Putat Jaya (Surabaya) | 65 m | — | | KOP. KEL. SUNGAI PANAS | Pekuburan Sei.Panas (Batam) | 51 m | — | **The honest verdict: _"dibangun di tanah kuburan"_ is not supported as a pattern.** Two cases are individually striking — a `Terverifikasi` land asset 481 m inside a private memorial park is a hard thing to explain — and they are worth chasing on imagery. But an existence claim needs the individual case to survive verification, and the rate carries no signal at all. ## Finding 3 — groundwork for [08](../README.md#08--exact-geometry-refinement-of-05-and-06) Distance to the nearest _pasar_ (`amenity=marketplace`), now measured exactly rather than in H3 bands: median **10.5 km**, with 689 KDMP within 250 m, 1,806 within 500 m and 9,716 within 2 km. The village-centre reference is in the same table for every cooperative. ## What this cannot settle **A coordinate in a field may never have been a building.** This is 04's ambiguity in a new place, and the evidence here leans _against_ the reassuring reading. If SIMKOPDES coordinates were desa centroids, a desa that is mostly rice fields would produce a point in the middle of a rice field automatically — and would land inside farmland far more often than an OSM village node does, which is exactly the 2.4× we measured. The sharpest available test is road distance, from [05](../05-road-access/): a building stands next to a road, a centroid need not. It does not clear them. | | Within ~140 m of a mapped road | | ------------------------------------------------- | ------------------------------ | | All KDMP, reweighted to the candidates' provinces | **72.5%** | | The 538 candidates | **59.5%** | **The candidates are _less_ roadside than comparable cooperatives, not more.** That is consistent with genuine placement out in the fields, and equally consistent with coordinates that were never buildings. 276 candidates are both roadside and confirmed cropland — 133 of them `Terverifikasi` — and that subset is where the centroid explanation is weakest, but "weakest" is not "excluded". **No aggregate from this report should be written as "N cooperatives were built in paddy fields."** The supportable sentence is _"N cooperatives are **recorded at** a location inside a paddy field"_ — and if the ministry answers that the coordinates are wrong, that is a different story, not a smaller one: it means the registry does not know where its own cooperatives are. ### Other limits - **The asymmetry rule, harder here than anywhere else.** OSM maps ~9,100 burial grounds against ~84,000 desa — roughly 10% coverage. A hit is strong evidence; a miss is _no_ evidence. Every "distance to nearest X" column is an **upper bound** on the truth. [`osm_landuse_coverage_by_province.csv`](osm_landuse_coverage_by_province.csv) is the check: Gorontalo has **no** KDMP with a mapped burial ground within 2 km, and the Papua provinces sit at 1–2%. Absence there is absence of mappers. - **The village-core distance is weak and is not used as a screen.** `place=village` is one node for a desa that may be kilometres across, so the "49.3% of KDMP are more than 800 m from any village node or place of worship" figure in the per-cooperative table mostly measures where mappers drop a label. It is in the output for context; do not build a claim on it. - **`landuse=farmland` is not `sawah`.** OSM Indonesia tags wet rice fields as farmland (`landuse=paddy` is effectively unused), so this cannot separate a paddy field from a dry field or a plantation edge. The WorldCover cross-check says "cropland", not "rice". - **Polygon quality is uneven even after filtering.** The 2-village threshold catches the worst tracings; a 117 km² "field" in Majalengka is still a large object to call one field, even with no village node inside it. - OSM changes daily; this is a 2026-08-07 snapshot. ## Outputs | File | Contents | | ------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | [`farmland_candidates.csv`](farmland_candidates.csv) | **538 rows** — the shortlist, with depth, polygon quality, WorldCover class, road distance, land status, `imagery_url` | | [`cemetery_candidates.csv`](cemetery_candidates.csv) | all 22 burial-ground hits | | [`farmland_funnel.csv`](farmland_funnel.csv) | the 2,209 → 538 filter chain | | [`null_comparison.csv`](null_comparison.csv) | KDMP vs village nodes vs minimarkets, per class | | [`landuse_pip_summary.csv`](landuse_pip_summary.csv) | national counts and distance bands per class | | [`osm_landuse_coverage_by_province.csv`](osm_landuse_coverage_by_province.csv) | **read before treating any absence as evidence** | | `kopdes_landuse_context.csv` | per-cooperative, all classes, joins to [03](../03-population-coverage/)/[05](../05-road-access/) on `cooperative_id` (gitignored, ~28 MB, rebuilds in ~1 min once the index is cached) |