ZaraatDost is a satellite-imagery training dataset for AI-powered parcel boundary segmentation, agricultural field-boundary extraction, and cadastral digitization research.

Dataset repository: AdilMunawar/Zaraatdost
Related model repository: AdilMunawar/Model15
Created and maintained by: Adil Munawar

This dataset is organized as multiple downloadable ZIP archives, each representing a geographic training subset. The repository also includes preview images showing RGB imagery, boundary masks, and overlay examples.

Repository Layout

The repository currently follows this practical structure:

Zaraatdost/
├── datasets/
│   ├── alam_arain_dataset.zip
│   ├── dino_mako_dataset.zip
│   ├── khansar_dataset.zip
│   ├── sindh_dataset.zip
│   └── agis_v2_monthly_2024_2025_backup_20260630_062415/
│
├── previews/
│   ├── alam_arain_preview.png
│   ├── dino_mako_preview.png
│   ├── khansar_preview.png
│   └── sindh_preview.png
│
└── README.md

Dataset Archives

The dataset is split into regional ZIP archives:

Archive	Description
`datasets/alam_arain_dataset.zip`	Parcel-boundary training data for the Alam Arain region.
`datasets/dino_mako_dataset.zip`	Parcel-boundary training data for the Dino Mako region.
`datasets/khansar_dataset.zip`	Parcel-boundary training data for the Khansar region.
`datasets/sindh_dataset.zip`	Parcel-boundary training data for the Sindh region.

Each ZIP archive is expected to contain paired satellite image patches and boundary masks suitable for binary semantic segmentation.

Recommended internal ZIP structure:

region_dataset.zip
├── train/
│   ├── images/
│   │   ├── sample_000001.png
│   │   ├── sample_000002.png
│   │   └── ...
│   └── masks/
│       ├── sample_000001.png
│       ├── sample_000002.png
│       └── ...
│
└── val/
    ├── images/
    │   ├── sample_000001.png
    │   ├── sample_000002.png
    │   └── ...
    └── masks/
        ├── sample_000001.png
        ├── sample_000002.png
        └── ...

If an archive contains a slightly different root folder name, extract the archive first and point the training script to the extracted dataset root.

Preview Gallery

The repository includes visual previews for each regional subset. Each preview image shows satellite imagery, corresponding parcel-boundary masks, and overlay examples.

Alam Arain Preview

Dino Mako Preview

Khansar Preview

Sindh Preview

What the Preview Images Show

The preview files are designed to help users visually understand the dataset format:

Top row:     RGB satellite imagery patches
Middle row:  Binary parcel-boundary masks
Bottom row:  Boundary overlays on satellite imagery

The masks contain thin line structures representing field, road-side, plot, or parcel separators. These masks are used as target labels for training segmentation models.

Dataset Purpose

The purpose of ZaraatDost is to train and improve deep-learning models that can automatically identify visible parcel and field boundaries from satellite imagery.

This dataset supports:

Parcel boundary segmentation
Cadastral digitization assistance
Agricultural field-boundary extraction
Boundary probability mask generation
Land-record modernization research
Human-in-the-loop GIS mapping workflows
Training HRNet-W48 + U-Net segmentation models

Task Definition

This dataset is designed for binary semantic segmentation.

Input:   RGB satellite image patch
Target:  Single-channel binary parcel boundary mask
Class 0: Background / non-boundary
Class 1: Parcel or field boundary

During training, masks are commonly loaded as grayscale PNGs and converted to binary tensors:

mask = cv2.imread(mask_path, cv2.IMREAD_GRAYSCALE)
mask = (mask > 127).astype(np.float32)

Image and Mask Format

RGB Satellite Images

Format:      PNG or JPG inside ZIP archives
Channels:    3-channel RGB
Typical size: 512 x 512 pixels
Content:     Satellite imagery patches

Boundary Masks

Format:      PNG inside ZIP archives
Channels:    Single-channel grayscale
Background:  0
Boundary:    255 recommended
Training:    Converted to 0.0 / 1.0

Download and Extract with Hugging Face Hub

Example for downloading one archive:

from huggingface_hub import hf_hub_download
import zipfile
from pathlib import Path

repo_id = "AdilMunawar/Zaraatdost"
archive_name = "datasets/alam_arain_dataset.zip"

zip_path = hf_hub_download(
    repo_id=repo_id,
    filename=archive_name,
    repo_type="dataset"
)

extract_dir = Path("./zaraatdost_alam_arain")
extract_dir.mkdir(parents=True, exist_ok=True)

with zipfile.ZipFile(zip_path, "r") as z:
    z.extractall(extract_dir)

print("Extracted to:", extract_dir)

Download all regional archives:

from huggingface_hub import hf_hub_download
from pathlib import Path
import zipfile

repo_id = "AdilMunawar/Zaraatdost"
archives = [
    "datasets/alam_arain_dataset.zip",
    "datasets/dino_mako_dataset.zip",
    "datasets/khansar_dataset.zip",
    "datasets/sindh_dataset.zip",
]

root = Path("./zaraatdost_extracted")
root.mkdir(parents=True, exist_ok=True)

for archive in archives:
    zip_path = hf_hub_download(
        repo_id=repo_id,
        filename=archive,
        repo_type="dataset"
    )

    region_name = Path(archive).stem
    out_dir = root / region_name
    out_dir.mkdir(parents=True, exist_ok=True)

    with zipfile.ZipFile(zip_path, "r") as z:
        z.extractall(out_dir)

    print("Extracted", archive, "to", out_dir)

Example PyTorch Dataset Loader

import glob
import cv2
import numpy as np
from torch.utils.data import Dataset

class BoundaryDataset(Dataset):
    def __init__(self, root_dir, split="train", transform=None):
        self.images = sorted(glob.glob(f"{root_dir}/{split}/images/*.png"))
        self.masks = sorted(glob.glob(f"{root_dir}/{split}/masks/*.png"))
        self.transform = transform

        assert len(self.images) == len(self.masks), (
            f"Mismatch: {len(self.images)} images vs {len(self.masks)} masks"
        )

    def __len__(self):
        return len(self.images)

    def __getitem__(self, idx):
        image = cv2.imread(self.images[idx])
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

        mask = cv2.imread(self.masks[idx], cv2.IMREAD_GRAYSCALE)
        mask = (mask > 127).astype(np.float32)

        if self.transform:
            aug = self.transform(image=image, mask=mask)
            image = aug["image"]
            mask = aug["mask"]

        return image, mask.unsqueeze(0)

Recommended Augmentations

Recommended training augmentations:

RandomRotate90
HorizontalFlip
VerticalFlip
Transpose
ShiftScaleRotate
RandomBrightnessContrast
HueSaturationValue
GaussNoise
MotionBlur
Normalize
ToTensorV2

These augmentations help the model generalize across different orientations, fields, crop patterns, brightness conditions, and image qualities.

Recommended Training Configuration

The dataset is suitable for HRNet-W48 + U-Net training.

ENCODER = "tu-hrnet_w48"
IN_CHANNELS = 3
NUM_CLASSES = 1
PATCH_SIZE = 512
BATCH_SIZE = 8
ACCUMULATION_STEPS = 2
MIXED_PRECISION = True
BCE_POS_WEIGHT = 4.0
METRIC_THRESHOLD = 0.35

Suggested combined loss:

Total Loss =
    0.30 * BCEWithLogitsLoss
  + 0.40 * DiceLoss
  + 0.20 * FocalLoss
  + 0.08 * Edge Loss
  + 0.02 * Continuity Loss

Expected Model Outputs

Models trained on this dataset can generate:

prediction_probability.tif
prediction_binary.tif
prediction_preview.png
confidence_stats.json

In production workflows, these outputs can be post-processed into:

clean_boundary_mask.tif
skeleton_repaired.tif
final_parcel_polygons.geojson
final_parcel_polygons.shp

Related Model

This dataset is designed to support the parcel boundary segmentation model:

AdilMunawar/Model15

Recommended architecture:

HRNet-W48 encoder + U-Net decoder

Intended Use

This dataset is intended for:

Training parcel boundary detection models
Fine-tuning satellite imagery segmentation models
Agricultural field boundary extraction
Cadastral and land-record digitization research
Human-in-the-loop GIS mapping workflows
Probability-mask based topology generation

Not Intended For

This dataset is not intended to be used as:

A legal cadastral ownership record
A replacement for official land surveys
A source for determining land ownership rights
A standalone legal boundary authority

Predicted outputs from models trained on this dataset should be reviewed by GIS professionals or domain experts before official use.

Dataset Limitations

Model quality learned from this dataset may depend on:

Satellite image resolution
Seasonal crop variation
Field visibility
Shadows, haze, cloud, or blur
Difference between training regions and inference regions
Quality and consistency of boundary masks
Boundaries that are not visually visible in imagery

Some legal cadastral boundaries may not appear in satellite imagery and therefore cannot be reliably learned from imagery alone.

Ethical and Practical Considerations

This dataset supports visual boundary detection, not land ownership determination.
Outputs should be treated as assistive mapping layers.
Human validation is recommended for land-record or cadastral workflows.
The dataset should be used responsibly according to applicable imagery, mapping, and data policies.

Suggested Citation

Adil Munawar — ZaraatDost Parcel Boundary Training Dataset for AI-based Cadastral and Agricultural Boundary Segmentation

Credit

Created and maintained by Adil Munawar.

Downloads last month: 37

Total file size:

19 GB

Models trained or fine-tuned on AdilMunawar/Zaraatdost

AdilMunawar/Model15

Image Segmentation • Updated 1 day ago