BaraaLazkani/Middle_East_AOD_Dataset

Middle East AOD Dataset

Paired Sentinel-2 L1C satellite imagery and AERONET ground-truth Aerosol Optical Depth (AOD) at 550nm, for 20 monitoring sites across the Middle East and Eastern Mediterranean (2015–2025).

Built to train models that estimate AOD directly from satellite imagery. Full pipeline code and downstream feature engineering / modeling live in the companion GitHub repo: DualMind-Lab/middle_east_AOD_estimation.

Dataset Structure

├── dataset.csv           # 6,183 labeled rows: AOD_550nm + metadata + image_path
└── images/
    ├── metadata.csv       # same 6,183 rows, paths relative to images/ — used by HF's dataset viewer
    └── {site_name}/
        └── {yyyymmdd_HHMMSS}.tif   # 13-band GeoTIFF, 10m resolution, ~500x500px, UINT16

images/ contains exactly the 6,183 files referenced by dataset.csv/metadata.csv — no unlabeled/orphan images are included.

Two CSVs exist because they serve different consumers:

• dataset.csv — original columns/paths, written for the middle_east_AOD_estimation pipeline (see below).
• images/metadata.csv — the Hugging Face imagefolder metadata convention: file_name relative to images/, plus the label/metadata columns. Without this file, the Hub's dataset viewer auto-infers image-classification labels from the site subfolder names, which is wrong for this dataset — the viewer reads this file instead.

dataset.csv columns

Column	Description
AERONET_Site	Ground station name
Date(dd:mm:yyyy), Time(hh:mm:ss)	Timestamp of the satellite overpass (UTC)
Day_of_Year, Day_of_Year(Fraction)	Seasonal position
AOD_550nm	Label. Mean of ≥2 AERONET measurements within ±30 min of the overpass
Site_Latitude(Degrees), Site_Longitude(Degrees), Site_Elevation(m)	Station metadata
Solar_Zenith_Angle(Degrees)	Sun angle at measurement time
image_path	Relative path to the matching image

Image bands (index → band → native resolution)

0	1	2	3	4	5	6	7	8	9	10	11	12
B01 (60m)	B02 (10m)	B03 (10m)	B04 (10m)	B05 (20m)	B06 (20m)	B07 (20m)	B08 (10m)	B8A (20m)	B09 (60m)	B10 (60m)	B11 (20m)	B12 (20m)

20m and 60m bands are upsampled to 10m via nearest-neighbour interpolation (no invented values); pixel values are UINT16, scaled ×10000 (L1C TOA reflectance).

Using this with middle_east_AOD_estimation

dataset.csv's image_path column is written relative to that repo's layout, e.g. Data_Collection/Dataset/images/AgiaMarina_Xyliatou/20151121_084033.tif. To use this dataset with the pipeline as-is (no path rewriting), download it into the repo at:

middle_east_AOD_estimation/
└── Data_Collection/
    └── Dataset/          ← this dataset's dataset.csv + images/ go here

If you're using the dataset standalone, join image_path on the images/ folder here and ignore the Data_Collection/Dataset/ prefix.

Data Collection Summary

1. AERONET measurements — downloaded globally (Version 3, Level 2.0, cloud-cleared), filtered to 20 Middle East / Eastern Mediterranean sites with usable coverage from 2015 onward (Sentinel-2 era), then AOD at 550nm was interpolated from the 500nm/675nm channels via the Ångström exponent.
2. Sentinel-2 imagery — for each AERONET measurement, the matching Sentinel-2 L1C scene within ±30 minutes was downloaded (5×5 km patch centered on the station, all 13 bands) and merged into a single 13-band GeoTIFF.
3. Matching & aggregation — each image was matched to its closest AERONET readings (±30 min); readings were averaged into one label per image, and images backed by fewer than 2 measurements were dropped. This follows the ±30 min / ≥2-measurement matchup protocol used in NASA's Dark Target validation and established in the AOD collocation literature (Sayer et al. 2020; Virtanen et al. 2018).

The full pipeline, including intermediate row counts, per-site statistics, and the source scripts, is documented in Data_Collection/README.md of the companion repo.

Dataset Summary

Samples	6,183
Sites	20
Countries	Cyprus, Egypt, Israel, Kuwait, Oman, Saudi Arabia, Turkey, UAE
Date range	2015-08-20 → 2025-11-30
Label	AOD at 550nm (mean of ≥2 AERONET measurements)

Sites

Site	Country	Samples	Site	Country	Samples
AgiaMarina_Xyliatou	Cyprus	465	Migal	Israel	338
CUT-TEPAK	Cyprus	831	SEDE_BOKER	Israel	383
Nicosia	Cyprus	493	Technion_Haifa_IL	Israel	155
Troodos_CAO	Cyprus	47	Weizmann_Institute	Israel	297
Cairo_EMA_2	Egypt	317	Kuwait_University	Kuwait	424
El_Farafra	Egypt	89	Shagaya_Park	Kuwait	240
Qena_SVU	Egypt	109	University_of_Nizwa	Oman	99
Eilat	Israel	155	KAUST_Campus	Saudi Arabia	560
IMS-METU-ERDEMLI	Turkey	515	DEWA_ResearchCentre	UAE	69
Masdar_Institute	UAE	376	Mezaira	UAE	221

Uses

Direct use: training/evaluating models that estimate AOD at 550nm from Sentinel-2 L1C imagery, and related remote-sensing regression research.

Out-of-scope use: this dataset is not validated for real-time or safety-critical air-quality monitoring or decision-making. It should not be assumed to generalize to:

• Regions outside the 20 sites/8 countries listed below — in particular, no AERONET stations exist for Lebanon, Syria, Qatar, Yemen, or Palestine, so the dataset has no coverage there despite being labeled "Middle East."
• Cloudy or heavily hazy conditions — AERONET Level 2.0 data is cloud-screened, so labels skew toward clear-sky scenes.
• Sites/seasons with little representation — coverage is heavily imbalanced toward Cyprus and Israel sites (e.g. CUT-TEPAK: 831 samples) versus sites like DEWA_ResearchCentre (69 samples); 2015 has only 37 samples since Sentinel-2 L1C only became available mid-year.

Limitations and Biases

• Geographic bias: sample density follows AERONET station placement, not population or pollution patterns — dense urban/industrial areas without a station are unrepresented.
• Temporal matching uncertainty: labels are the mean of AERONET readings within ±30 minutes of the satellite overpass, following the NASA Dark Target validation protocol; this introduces some label noise relative to the exact overpass instant (see Sayer et al. 2020 for quantification, σ ≈ 0.019 at 30 min).
• Sensor drift/calibration: AERONET instruments are periodically recalibrated across sites and years; no per-instrument calibration metadata is included.

License

MIT (this repository's code/labels/organization). Underlying source data has its own terms:

• AERONET measurements are provided by NASA GSFC under their data usage policy — cite Holben et al. (1998) when publishing results.
• Sentinel-2 imagery is provided by the Copernicus Programme (ESA) under the Copernicus open data policy — include the attribution "Contains modified Copernicus Sentinel data [year]."

Citation

@misc{lazkani2026middleeastaod,
  author = {Lazkani, Baraa and Ibrahim, Modar},
  title = {Middle East AOD Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/BaraaLazkani/Middle_East_AOD_Dataset}
}

Also cite the underlying sources per their policies:

• Holben, B.N., et al. (1998). AERONET — A federated instrument network and data archive for aerosol characterization. Remote Sensing of Environment.
• "Contains modified Copernicus Sentinel data [year], processed by ESA."

Dataset Card Contact

Baraa Lazkani and Modar Ibrahim — lazkani.baraa.official@gmail.com · github.com/DualMind-Lab/middle_east_AOD_estimation