HERITAGE Dataset

A satellite image dataset for monitoring archaeological sites, derived from Planet Labs monthly basemap mosaics at 4.77 m/pixel resolution. HERITAGE covers 1,982 sites across 16 countries with monthly observations from January 2016 through May 2025.

The Afghanistan subset contains 1,943 sites with binary looting labels (898 looted, 1,045 preserved), per-site bounding masks, and month-of-disturbance annotations for 118 sites with confirmed change events. The global subset adds 39 sites across 15 additional countries on five continents. Among public archaeological monitoring datasets, HERITAGE is larger than prior releases on number of sites (1,982), countries (16), images (212,776), and months of coverage (113). Site coordinates are withheld to protect the locations from exploitation; the processing pipeline and ground-truth labels are released with the imagery.

Geographic coverage

Geographic distribution of HERITAGE sites. (a) World map of the 40 global monitoring locations across 16 countries: 39 individual archaeological sites in 15 countries (blue circles) and the Afghanistan region (red star, containing 1,943 sub-sites). (b) Afghanistan detail showing the spatial distribution of 898 looted (orange) and 1,045 preserved (blue) archaeological sites.

The 39 global sites span 15 countries: Belize (1), Cambodia (1), Ecuador (2), Egypt (5), Italy (5), Mali (1), Pakistan (2), Peru (6), Sudan (1), Sweden (2), Syria (2), Thailand (2), Turkey (3), Ukraine (1), and the USA (5). Each global site has 100-101 monthly observations. Geographic metadata is stripped from the released PNG files; coordinates are not distributed with the imagery.

Sample imagery

Sample RGB composites from 12 HERITAGE sites spanning 10 countries. Images are displayed at 186 x 186 pixels (rescaled for global sites). Each image shows a single monthly observation from the middle of the site's time series.

Dataset summary

The dataset is partitioned into an Afghanistan subset (1,943 fully annotated sites) and a global subset (39 sites across 15 countries, imagery only). Subset-specific attributes are shown side by side; shared attributes are listed below the divider. Combined totals are 1,982 sites, 212,776 images, and 113 distinct months of monthly coverage from January 2016 through May 2025.

Attribute	Afghanistan sites	Global sites
Number of sites	1,943 (898 looted, 1,045 preserved)	39 across 15 countries
Number of images	~210,000	~3,900
Temporal range	January 2016 to December 2024	January 2017 to May 2025
Months per site	up to 108 (median 107)	100-101
Image dimensions	186 x 186 pixels	per-site majority dimension
Site footprint	~1 km x 1 km (fixed)	0.1 - 50.6 km^2
Binary site mask	Provided (1,943 PNGs)	Not provided
Looting label	Provided (898 looted, 1,045 preserved)	Not provided
Change-month label	Provided (118/898 confirmed)	Not provided

Shared across both subsets: 3 spectral bands (R, G, B) plus a 1-channel data validity mask; 4.77 m/pixel spatial resolution (Web Mercator zoom level 15); monthly cadence; 4-band PNG (RGBA, 8-bit unsigned per channel) image format; single-band PNG (binary, 0/255) mask format; Planet Labs monthly basemap mosaics (PlanetScope, Dove constellation) as the data source.

Dataset statistics. (a) Distribution of looted vs. preserved labels in the Afghanistan subset. (b) Temporal distribution of confirmed looting events by year (118 sites with known change month, peak in 2019). (c) Number of monitoring sites per country (log scale); Afghanistan accounts for 1,943 of the 1,982 sites.

Comparison with prior datasets

The table below compares HERITAGE to earlier archaeological remote-sensing datasets. Among the publicly released datasets in this list, only DAFA-LS (Vincent et al., 2024) predates HERITAGE; HERITAGE adds multi-country coverage, more months of observation, and per-site change-month labels. Image counts are reported only where the source paper specifies them; "---" indicates not reported. "Change month" is a per-site month-of-disturbance label for looted sites. "Site mask" is a binary raster delineating the archaeological area within each image chip.

Dataset	Sites	Countries	Months	Images	Change month	Site mask	Public
Casana (2015)	14	1	1	---	---	---	No
Parcak et al. (2016)	200+	1	2-4	---	---	---	No
Tapete & Cigna (2016)	1	1	---	---	---	---	No
Lauricella et al. (2017)	1	1	1	---	---	---	No
Tadesse et al. (2026a)	1,943	1	96	---	---	---	No
Tadesse et al. (2026b)	1,943	5	96	---	Yes	---	No
Vincent et al. (2024) [DAFA-LS]	675	1	96	55,480	---	Yes	Yes
HERITAGE	1,982	16	113	212,776	Yes	Yes	Yes

Directory structure

HERITAGE/
  dataset/
    ground_truth.csv
    Afghanistan/
      looted_0/
        2016_01.png
        2016_02.png
        ...
        mask.png
      looted_1/
        ...
      preserved_0/
        ...
      ...
    Belize_Lubaantun/
      2017_01.png
      ...
    Cambodia_Panteay_Chamar/
      ...
    [37 additional global site directories]

Layout notes

• ground_truth.csv lists the three label fields for each of the 1,943 Afghanistan sites: site_name, looted (binary), and looted_month (integer month index when looting was detected; -1 if confirmed but month unknown; 0 for preserved sites).
• Afghanistan site directories follow the naming pattern {looted,preserved}_N, where N is a zero-indexed site identifier. Each contains monthly RGBA PNG chips and a mask.png raster delineating the archaeological area.
• Global site directories follow the naming pattern Country_SiteName (39 directories across 15 countries). They contain monthly RGBA PNG chips; no per-site masks are provided.
• File names follow YYYY_MM.png, where YYYY is the four-digit year and MM is the two-digit month.
• Images are stored as four-channel PNGs (height x width x 4): the first three channels are R, G, B spectral values; the fourth (alpha) channel is a binary data-validity mask (255 for valid pixels, 0 for zero-padding outside the site extent).
• Geographic metadata (coordinate reference system, affine transform) is removed during conversion to prevent direct geolocation of sites from the image files.

Ground-truth labels

The schema below describes the three fields in dataset/ground_truth.csv, provided for each of the 1,943 Afghanistan sites. Labels were assigned by archaeological experts at ICONEM through visual interpretation of high-resolution satellite imagery, corroborated by field survey data where available. The annotation protocol first classified each site as looted or preserved based on the presence of surface disturbance in the imagery, then identified the specific month of disturbance for looted sites by inspecting sequential monthly images for the first appearance of pitting or trenching.

Field	Type	Description
site_name	string	Identifier in the format {looted,preserved}_N.
looted	binary	1 = looted, 0 = preserved.
looted_month	integer	Month index when looting was detected; -1 if confirmed but month unknown; 0 for preserved sites.

Looting sequence examples

Temporal progression around looting events at three Afghanistan sites. Each row shows six consecutive monthly RGB composites. The change month (red label, red border) marks the first observation of visible surface disturbance. Preceding months show the undisturbed state; subsequent months show post-looting conditions.

Site bounding masks

Binary bounding masks are provided for the Afghanistan subset only. Each Afghanistan site directory contains a mask.png file where foreground pixels (value 255) indicate the archaeological area and background pixels (value 0) indicate areas outside the site boundary. Masks are 186 x 186 pixels and aligned to the image chip grid. They were produced by manually digitising the polygon of each site against high-resolution reference imagery, then rasterising. All 1,943 Afghanistan sites (898 looted, 1,045 preserved) include a mask. No masks are provided for global sites.

Binary bounding mask examples for six randomly selected Afghanistan sites with full pixel coverage. Top row: RGB composites from the middle of each site's time series. Bottom row: the same images with the bounding mask overlaid in red, delineating the archaeological area within the 186 x 186 pixel chip. Left three columns show looted sites; right three columns show preserved sites. Masks vary in size and shape, reflecting the irregular boundaries of the underlying archaeological features.

Site spatial extents

Global site footprints range from 0.1 km^2 (Pakistan, Charsadda NW; Belize, Lubaantun) to 50.6 km^2 (Turkey, Bintepe North). Afghanistan sites each cover an approximately 1 km x 1 km footprint (186 x 186 pixels at 4.77 m/pixel).

Spatial footprint of each HERITAGE site in km^2. Global sites (blue) range from 0.1 to 50.6 km^2. Afghanistan sites (orange) each cover an approximately 1 km x 1 km footprint.

Land cover context

Land cover composition around each site was extracted from the ESA WorldCover 2021 product (10 m resolution, 11 classes derived from Sentinel-1 and Sentinel-2). Afghanistan sites are dominated by bare or sparse vegetation (61.5%) and cropland (27.7%), consistent with the arid steppe and irrigated agriculture of the northern Afghan provinces where sites concentrate. Global sites span a wider range: Belize (Lubaantun) sits within dense tropical forest (99.6% tree cover); Sudan (Uronarti), a Nile island fortress, is surrounded by permanent water (69.8%); Italian sites occupy Mediterranean grassland and forest; and Thai sites show mixed cropland, forest, and built-up areas. This diversity is relevant for transfer-learning experiments, as models trained on Afghan sites must generalize across different spectral backgrounds.

Mean percentage of pixels per WorldCover class within each site's spatial extent, aggregated by country:

Country	N	Tree	Shrub	Grass	Crop	Built	Bare	Water	Wetland	Other
Afghanistan	1943	0.5	3.5	3.9	27.7	2.5	61.5	0.3	0.0	0.0
Belize	1	99.6	0.0	0.4	0.0	0.0	0.0	0.0	0.0	0.0
Cambodia	1	21.5	0.0	15.7	54.9	7.0	0.4	0.6	0.0	0.0
Ecuador	2	91.6	0.1	3.9	0.1	0.0	0.8	3.5	0.0	0.0
Egypt	5	4.3	1.6	1.1	8.1	9.3	75.6	0.0	0.1	0.0
Italy	5	29.7	3.2	44.5	11.5	1.7	0.8	8.6	0.0	0.0
Mali	1	0.0	14.6	66.6	11.3	0.0	0.4	0.0	7.0	0.0
Pakistan	2	37.4	25.1	4.1	18.4	5.2	9.8	0.0	0.0	0.0
Peru	6	24.2	3.0	36.1	10.5	3.7	20.2	1.9	0.4	0.0
Sudan	1	1.2	0.9	0.0	0.0	0.0	28.0	69.8	0.0	0.0
Sweden	2	19.6	0.0	72.2	4.1	0.2	1.4	2.4	0.1	0.0
Syria	2	0.2	0.6	14.9	36.2	3.5	40.6	4.0	0.0	0.0
Thailand	2	32.2	0.0	2.5	34.6	29.3	0.0	1.4	0.0	0.0
Turkey	3	22.5	0.1	36.7	39.6	0.9	0.2	0.0	0.0	0.0
USA	5	38.4	0.0	49.3	5.9	2.2	0.0	4.1	0.0	0.0
Ukraine	1	10.0	0.0	36.1	1.5	28.6	0.4	23.4	0.0	0.0

Land cover composition by country, derived from ESA WorldCover 2021. Each bar shows the mean percentage of land cover classes across sites in that country. Site counts are shown in parentheses.

Spectral characteristics

Per-band pixel value statistics for the Afghanistan subset (non-padding pixels, 30 sampled sites). Values are 8-bit unsigned integers in the range 0-255.

Band	Mean	Std
Red	173.0	17.9
Green	151.4	15.1
Blue	115.1	14.6

Mean values for all three spectral bands decreased between 2016 and 2019, then stabilized from 2020 onward. This radiometric shift reflects evolving sensor calibration and compositing methods in the Planet basemap product over the observation period. Users training models across years should account for these temporal shifts through normalization or domain adaptation. Across 30 sampled Afghanistan sites the mean no-data pixel fraction (pixels where all channels equal zero) is 4.7%, with a range of 0.0-35.1%. Global sites have higher no-data fractions (up to 66% for irregularly shaped sites such as Ecuador, Upano Sangay) because rectangular image chips contain padding around non-rectangular site footprints.

Radiometric trends across the HERITAGE dataset. (a) Yearly mean pixel values per band for Afghanistan (20 sampled sites), showing a decrease from 2016 to 2019 followed by stabilization. (b) Seasonal cycle for Afghanistan, with summer peaks reflecting arid-zone solar geometry. (c) Cross-site yearly Red band trends for four sites in different climatic zones. (d) Cross-site seasonal cycles showing contrasting phenological patterns across geographies.

Image acquisition and processing

All imagery is sourced from Planet Labs monthly basemap mosaics, which composite daily PlanetScope captures (Dove constellation, 3.0-3.7 m native ground sample distance) into cloud-free monthly products using a best-pixel selection strategy. Mosaics are delivered at Web Mercator zoom level 15, which corresponds to 4.77 m/pixel at the equator; effective ground pixel size varies with latitude (about 4.0 m at Afghanistan's latitude of ~33 deg N). For Afghanistan sites, imagery was clipped to 1 km x 1 km areas centred on each site's coordinates. For global sites, the corresponding site polygons were used to extract the imagery. The conversion pipeline reads each four-band GeoTIFF and writes a PNG of 186 x 186 pixels for Afghanistan, or the most common dimension across monthly images for each global site. Coordinate reference system and affine transform are removed during conversion.

The combined dataset spans 113 months from January 2016 through May 2025: up to 108 monthly observations per Afghanistan site (median 107) and 100-101 per global site (37 of 39 have 101 images, 2 have 100). Coverage gaps occur where Planet Labs mosaics were unavailable for a given month and location. No interpolation or gap-filling was performed; missing months are absent from the directory listing.

Label validation

Looting labels were produced by archaeological experts at ICONEM. Multiple analysts reviewed each site; disagreements were resolved through discussion. Of 898 looted sites, 118 have a confirmed change month; the remaining 780 are labelled as looted without a specific month (looted_month = -1). Two prior studies provide indirect validation: Tadesse et al. (2026a) trained classifiers on these labels and reported performance consistent with accurate annotations, and Tadesse et al. (2026b) used the temporal annotations to train and evaluate change detection models with results that corroborate the annotated change months.

Counts

Subset	Sites	Countries	Images	Temporal span
Afghanistan	1,943	1	~210,000	2016-01 to 2025-05 (up to 108 mo)
Global	39	15	~3,900	2016-01 to 2025-05 (100-101 mo)
Total	1,982	16	212,776	113 months

Loading

Load images with any standard image library that supports four-channel PNGs. Load all four channels explicitly and avoid loaders that apply alpha premultiplication, which would corrupt the RGB values.

Baseline results

Binary looting classification baselines on the Afghanistan subset, reported by Tadesse et al. (2026a). Results are averaged over 5-fold cross-validation with an 80/10/10 train/validation/test split. Feature-based methods use temporal mean aggregation; CNN methods use single-date imagery.

Method	Model	Accuracy	F1	AUC
Feature-based classifiers
Handcrafted	Logistic Regression	0.720	0.705	0.768
Handcrafted	Random Forest	0.716	0.693	0.783
SatCLIP embeddings	Logistic Regression	0.710	0.685	0.779
SatMAE embeddings	Random Forest	0.678	0.635	0.741
End-to-end CNNs
Single-date	ResNet-18	0.804	0.807	0.859
Single-date	EfficientNet-B0	0.870	0.862	0.927
Single-date	EfficientNet-B1	0.941	0.936	0.975

Repository contents

This repository ships the dataset documentation and the scripts that produce it; the imagery itself is hosted separately (see Download below).

.
|-- README.md                       this file
|-- LICENSE                         MIT license
|-- SECURITY.md                     vulnerability-disclosure policy
|-- CONTRIBUTING.md                 contributor license agreement notice
|-- THIRD_PARTY_NOTICES.md          third-party dependency attributions
|-- requirements.txt                pinned Python dependencies
|-- .gitignore
|-- figs/                           figures referenced from the README
|-- src/
|   |-- download_planet_mosaics.py  download monthly basemaps from Planet
|   |-- tif_to_png_4band.py         convert 4-band GeoTIFFs to RGBA PNGs
|   |-- generate_site_masks.py      rasterize site polygons to mask.png
|   `-- generate_lulc_analysis.py   reproduce the land-cover table and figure
`-- dataset/                        (not in this repo; download separately)

Installation

The scripts in src/ target Python 3.10 or newer. Pick one environment-setup path below, then install the pinned dependencies from requirements.txt.

Option A: conda

conda create -n heritage python=3.10
conda activate heritage
pip install -r requirements.txt

Option B: venv

python -m venv .venv
source .venv/bin/activate           # Linux / macOS
.venv\Scripts\activate              # Windows PowerShell
pip install -r requirements.txt

Planet API key

src/download_planet_mosaics.py reads the Planet API key from the PLANET_API_KEY environment variable. Obtain a key from https://www.planet.com/account/ and export it in the active shell:

export PLANET_API_KEY=<your-key>    # Linux / macOS
$env:PLANET_API_KEY = "<your-key>"  # Windows PowerShell

The key is read from the environment and never written to disk by these scripts.

Reproducing the dataset

1. Run src/download_planet_mosaics.py with --mode point and a CSV of site (name, latitude, longitude) rows for the Afghanistan-style 1 km x 1 km windows, or --mode polygon with a GeoJSON of site polygons for the global subset. Requires PLANET_API_KEY to be set (see Installation above).
2. Run src/tif_to_png_4band.py --mode all to convert the downloaded 4-band GeoTIFFs into the RGBA PNG layout described in Layout notes.
3. Run src/generate_site_masks.py --sites sites.csv --output dataset/Afghanistan with a CSV of (site_name, latitude, longitude, polygon_wkt) rows to produce the per-site mask.png files at 186 x 186 pixels and 4.77 m/pixel, centred on each site's coordinates.
4. Run src/generate_lulc_analysis.py to regenerate figs/fig_lulc_breakdown.png and the per-country land-cover table. The first invocation runs the fetch phase, which queries ESA WorldCover 2021 via the Microsoft Planetary Computer for each site and writes a local lulc_data.json cache; subsequent invocations reuse the cache via --phase analyze. The fetch phase needs the same site coordinates as step 1.

Site coordinates and polygons are withheld from the public release to protect sites from exploitation; contact the authors if access is required for research.

Companion analysis code

• Change detection: https://github.com/microsoft/WATCH
• Looting classification: https://github.com/microsoft/looted_site_detection

Both companion repositories are released under the MIT License. The dataset DOI and full citation will be assigned on publication of the data paper, HERITAGE: A Global Multi-Temporal Satellite Dataset for Archaeological Site Monitoring.