m0hamadk/paranal-seeing-dataset

---

license_name: eso-archive-terms license_link: http://archive.eso.org task_categories: - tabular-regression tags: - astronomy - atmospheric-science - weather - tabular - time-series - physics size_categories: - 10K<n<100K pretty_name: ESO Paranal Real Seeing + Site Instrument Dataset (2016-2026)

ESO Paranal Observatory - Real Seeing + Site Instrument Dataset (2016-2026)

Dataset Summary

Hourly, real-world atmospheric and astronomical-seeing measurements from the European Southern Observatory's Paranal site in Chile (24.6272 S, 70.4039 W, 2635 m a.s.l.), spanning April 2016 - July 2026 (36,706 rows). Built by merging three real ESO site-monitoring instruments (DIMM, LHATPRO, and a 30 m meteorological tower) on a common hourly timestamp. Every row is a real, instrument-measured observation - not simulated or forecast data.

Originally built to train a model predicting astronomical seeing from real atmospheric conditions (see the companion model on the Hub: search "paranal-seeing-xgboost").

Supported Tasks

• Tabular regression: predict seeing (real DIMM-measured astronomical seeing, in arcsec) from the 17 atmospheric feature columns.

Dataset Structure

Data Fields

Field	Description	Units
datetime	Hourly UTC timestamp	ISO 8601
airmass	Telescope airmass at time of DIMM measurement	dimensionless
pressure	Surface air pressure	hPa
temp_2m / temp_30m / temp_ground	Air temperature at 2 m / 30 m / ground	°C
temp_gradient_ground_30m	temp_ground - temp_30m	°C
dew_point_depression_2m	temp_2m - dew_point_2m	°C
wind_speed_10m / wind_speed_30m	Horizontal wind speed at 10 m / 30 m	m/s
wind_shear_10_30	wind_speed_30m - wind_speed_10m	m/s
wind_dir_10m	Wind direction at 10 m (hourly circular mean)	degrees
wind_w_std	Std. dev. of vertical wind speed within the hour (turbulence proxy)	m/s
rain_intensity	Rain intensity	%
ir_temperature	LHATPRO sky infrared temperature	°C
liquid_water_path	LHATPRO liquid water path	g/m²
pwv	Precipitable water vapour	mm
seeing	Target. Real DIMM-measured astronomical seeing	arcsec

Data Splits

No predefined split is included. For time-series-correct evaluation, do not split randomly by row - consecutive hours are highly autocorrelated. Instead, partition by contiguous time blocks (e.g. by ISO week) and hold out the most recent ~15% of blocks as a test set. See the companion GitHub repo for a reference implementation.

Dataset Creation

Source Data

All raw data are from the public ESO Science Archive ambient-conditions database for Paranal Observatory.

Data Collection and Processing

Each instrument logs at a different, irregular native cadence (roughly 1-2 minutes). Each source was independently aggregated to hourly statistics (mean for most quantities; circular mean for wind direction; standard deviation for vertical wind speed) and inner-joined on the shared hourly timestamp.

A fourth instrument, MASS (free-atmosphere turbulence), was deliberately excluded: its outputs are real-time turbulence measurements co-derived with DIMM seeing, so using them as model inputs would be target leakage.

Usage

from datasets import load_dataset

ds = load_dataset("csv", data_files="paranal_specific_dataset.csv")
print(ds["train"][0])

Limitations

Several of the strongest predictive features are hyper-local to Paranal's own instrumentation and are not reproducible from any general weather- forecast model. Paranal's own ambient database itself updates once per local night rather than in real time, which limits this dataset's use for live/real-time applications.

Licensing / Citation

Data sourced from ESO's public archive for research/educational use. Please credit the European Southern Observatory and the original archive (http://archive.eso.org) if you use this dataset.