Zaraatdost — Cadastral Parcel Boundary Segmentation

Paired satellite imagery and cadastral parcel-boundary masks for semantic segmentation of agricultural field boundaries in Pakistan. Each sample is a 512x512 RGB patch with a matching binary mask marking surveyed parcel edges.

At a glance

Total patches	277,297
Total size	78.3 GB
Patch size	512 x 512 px, 3-channel RGB
Image format	PNG, lossless
Mask format	PNG, binary (0 / 255)
Shards	131 WebDataset .tar
Source imagery	Google satellite basemap, zoom 19
Ground resolution	~0.30 m/px at the equator
Projection	EPSG:3857 (Web Mercator)
Boundary width	~5 px (2 px buffer either side of the surveyed line)

Splits

Split	Patches	Share	Shards	Size
train	235,518	84.9%	91	66.6 GB
val	29,743	10.7%	23	8.4 GB
test	12,036	4.3%	17	3.3 GB
total	277,297	100%	131	78.3 GB

Split methodology

Patches are extracted on a sliding window with a stride smaller than the patch size, so neighbouring patches overlap heavily. Assigning splits per patch would leak training pixels into validation and inflate every metric.

Instead, splits are assigned spatially:

1. Each mosaic is divided into 4096 x 4096 px regions.
2. Every region is deterministically hashed to one split.
3. A 128 px guard margin is stripped from each region edge, and any patch straddling a region boundary is discarded.

The result is that no patch in one split shares a single pixel with a patch in another, and the closest cross-split pair is separated by a 256 px gap. This was verified exhaustively over a 20,000 x 20,000 px grid: zero overlapping pairs. About 25% of candidate patches are dropped to buy this guarantee.

For a stricter test of generalisation, hold out entire mauzas rather than relying on the region split.

Label statistics

Measured on 360 randomly sampled patches:

Metric	Value
Mean boundary pixels per patch	4.41%
Median boundary pixels per patch	2.85%
Max boundary pixels in a patch	12.65%
Patches with no boundary	2.2%
Mean image size	262 KB
Mean mask size	3.0 KB
Mask values	[0, 255]

The classes are heavily imbalanced — roughly 4.4% positive. Plain BCE will collapse to predicting all-background. Use Dice, Focal, or clDice, or a weighted BCE, and report boundary-F1 with a tolerance rather than pixel accuracy.

Coverage by mauza

Mauza	Train	Val	Test	Total
kishtawar_merged_0105_381951_parcels__kacha_chohan	23,816	6,212	772	30,800
kishtawar_merged_0105_381951_parcels__kacha_mianwali_east	18,837	1,848	834	21,519
kishtawar_merged_0105_381951_parcels__kupra	16,412	2,677	276	19,365
kishtawar_merged_0105_381951_parcels__kacha_mianwali_west	16,671	2,024	0	18,695
kishtawar_merged_0105_381951_parcels__chankda	14,736	2,465	273	17,474
kishtawar_merged_0105_381951_parcels__chak_belay_shah	13,883	1,709	1,681	17,273
kishtawar_merged_0105_381951_parcels__bambli_khair_pur	15,273	358	729	16,360
kishtawar_merged_0105_381951_parcels__kacha_gopang	6,532	1,995	486	9,013
kishtawar_merged_0105_381951_parcels__sabzani_1	5,676	1,660	729	8,065
kishtawar_merged_0105_381951_parcels__dera_bhai	6,090	0	1,627	7,717
kishtawar_merged_0105_381951_parcels__mohari	6,632	496	298	7,426
kishtawar_merged_0105_381951_parcels__kacha_paru_shah	6,466	0	184	6,650
kishtawar_merged_0105_381951_parcels__bait_allah_wasaya	6,317	193	0	6,510
kishtawar_merged_0105_381951_parcels__rakh_kacha_shair_mehar	5,155	1,072	0	6,227
kishtawar_merged_0105_381951_parcels__khalti	5,592	0	0	5,592
kishtawar_merged_0105_381951_parcels__kacha_mahazi	4,633	838	0	5,471
kishtawar_merged_0105_381951_parcels__lut	4,486	668	244	5,398
kishtawar_merged_0105_381951_parcels__faiz_abad	5,200	29	0	5,229
kishtawar_merged_0105_381951_parcels__gharkan	4,352	300	196	4,848
kishtawar_merged_0105_381951_parcels__sabzan_pakka	4,056	519	268	4,843
kishtawar_merged_0105_381951_parcels__chack_veeha	4,469	0	330	4,799
kishtawar_merged_0105_381951_parcels__bait_dur_muhammad	2,859	793	701	4,353
kishtawar_merged_0105_381951_parcels__sabzani_tukra_2	2,216	2,136	0	4,352
kishtawar_merged_0105_381951_parcels__haji_pur	4,136	0	0	4,136
kishtawar_merged_0105_381951_parcels__nawaz_pur	3,458	0	617	4,075
kishtawar_merged_0105_381951_parcels__rakh_kacha_chohan	3,725	0	0	3,725
kishtawar_merged_0105_381951_parcels__kachi_aslam	2,181	0	863	3,044
kishtawar_merged_0105_381951_parcels__thul_nawab	2,753	248	0	3,001
kishtawar_merged_0105_381951_parcels__dahirwali	2,065	237	632	2,934
kishtawar_merged_0105_381951_parcels__fateh_pur_tewana	2,220	544	0	2,764
kishtawar_merged_0105_381951_parcels__kachi_zaman	2,744	0	0	2,744
kishtawar_merged_0105_381951_parcels__muslim_abad	2,546	0	0	2,546
kishtawar_merged_0105_381951_parcels__pir_bakhash	1,745	0	0	1,745
kishtawar_merged_0105_381951_parcels__bolani	1,673	67	0	1,740
kishtawar_merged_0105_381951_parcels__hamid_pur	1,411	0	0	1,411
kishtawar_merged_0105_381951_parcels__kot_ghulam_miran_shah	1,295	0	0	1,295
kishtawar_merged_0105_381951_parcels__kot_karam_khan	417	520	0	937
kishtawar_merged_0105_381951_parcels__kareema_khore_1	572	0	296	868
kishtawar_merged_0105_381951_parcels__manak	807	0	0	807
kishtawar_merged_0105_381951_parcels__bait_mir_ahmed	608	0	0	608

4 further mauzas omitted.

Structure

data/
  train/
    train-00000.tar
    train-00001.tar   (91 shards)
  val/
    val-00000.tar
    val-00001.tar   (23 shards)
  test/
    test-00000.tar
    test-00001.tar   (17 shards)
build_state.json

Each .tar is a WebDataset shard holding paired members:

<mauza>_<y>_<x>.png        RGB patch
<mauza>_<y>_<x>.mask.png   binary boundary mask

The key encodes the source mauza and the patch's pixel origin in that mosaic, so patches can be traced back to their location or regrouped spatially.

Usage

import webdataset as wds

base = "https://huggingface.co/datasets/AdilMunawar/Zaraatdost/resolve/main"
url = f"{base}/data/train/train-{{00000..00090}}.tar"

ds = (
    wds.WebDataset(url, shardshuffle=True)
    .decode("rgb8")
    .to_tuple("png", "mask.png")
    .shuffle(1000)
)

for img, mask in ds:
    pass

Or with datasets:

from datasets import load_dataset
ds = load_dataset("AdilMunawar/Zaraatdost", split="train", streaming=True)

Streaming is recommended — the full dataset is 78.3 GB and does not need to land on disk.

Construction

1. Cadastral parcel polygons are read from a GeoPackage and grouped by mauza.
2. For each mauza, satellite tiles covering the parcel extent are fetched at zoom 19 and mosaicked into a Web Mercator GeoTIFF.
3. Parcel boundaries are buffered by ~2 px and rasterised against the mosaic grid, producing a pixel-aligned binary mask.
4. The mosaic is cut into overlapping patches. Patches that are mostly black, uniform, or straddle a split region are discarded; most boundary-free patches are dropped, with a fraction retained as negatives.
5. Patches are written to WebDataset shards and uploaded.

The area of interest is eroded slightly before tile selection. Neighbouring villages that fall inside the bounding box but were never digitised would otherwise appear as imagery with no boundary labels, teaching a model that field edges of that texture are unlabelled.

Limitations

• Unlabelled neighbours. AOI erosion reduces but does not eliminate undigitised parcels appearing in patches near the edge of a mauza. Some false-negative label noise remains.
• Boundary width is a modelling choice. Masks are dilated lines, not the zero-width survey geometry. A model matching the labels exactly is matching a ~5 px band, so evaluate with a tolerance-based boundary F1, not raw IoU.
• Imagery is lossy at source. Basemap tiles arrive JPEG-compressed. Patches are stored losslessly to avoid a second generation of artefacts along the high-contrast field edges, but the source compression is already baked in.
• Temporal mismatch. Cadastral records and the basemap imagery were not captured at the same time. Recently subdivided or merged parcels may disagree with what is visible.
• Geographic scope. Limited to the surveyed mauzas listed above; generalisation to other regions, crop systems, or field geometries is untested.
• Overlapping patches. The sliding window means patches within a split overlap substantially. Effective sample size is well below the patch count.

Licensing and source imagery

The annotations in this dataset are released under cc-by-nc-4.0.

The imagery is derived from a third-party satellite basemap and remains subject to that provider's terms of service. It is included here for research and non-commercial use. Anyone intending commercial use, redistribution, or deployment should re-source the imagery from a provider whose licence permits it — Esri World Imagery, Bing Maps, Mapbox, or Sentinel-2/Planet for lower-resolution alternatives — and rebuild the patches against the same polygons.

Cadastral polygons derive from official land records. Verify your own rights to redistribute derivatives before reusing them.

Citation

@misc{zaraatdost_boundaries,
  title  = {Zaraatdost — Cadastral Parcel Boundary Segmentation},
  author = {Munawar, Adil},
  year   = {2026},
  url    = {https://huggingface.co/datasets/AdilMunawar/Zaraatdost}
}

---

Card generated 2026-09-02 from repository contents.