Paranal Seeing Prediction - XGBoost
Model Description
An XGBoost regression model that predicts real astronomical seeing (a
measure of atmospheric turbulence that limits telescope image quality) at
ESO's Paranal Observatory, trained on real, site-specific instrument data
(2016-2026) rather than generic weather-forecast features. Hyperparameters
were tuned with a 9-parameter Optuna search, validated on a held-out block of
weeks distinct from both training and final test data.
The model predicts log(1 + seeing); invert with np.expm1(prediction) to
recover seeing in arcsec.
An ensemble of this model with a companion neural network was also tested
and improved R\u00b2 by only 0.0025 (within noise) - not enough to justify a
two-model production system, so this XGBoost model alone is the recommended
model.
Intended Uses & Limitations
Intended use: retrospective / statistical analysis of the relationship
between real atmospheric conditions and astronomical seeing at Paranal (e.g.
site-characterization research, as a reference/starting point for building
an equivalent model at another observatory with comparable instrumentation).
Not intended for: live, forward-looking seeing forecasting. Several of
the strongest input features (local meteo-tower and radiometer readings) are
hyper-local to Paranal's own instrumentation and are not produced by any
general weather-forecast model; Paranal's own ambient database itself
updates once per local night rather than in real time. A deployable
forecasting tool would need to be re-trained on forecast-compatible inputs
(e.g. reanalysis pressure-level data), trading some accuracy for lead-time
forecastability.
Out-of-distribution use: this model is specific to Paranal's climate and
altitude (2635 m, Atacama desert). Applying it to another site's conditions
without retraining is not recommended - see feature_ranges.json (included)
for the valid input range seen during training; inputs far outside these
ranges should not be trusted.
Training Data
Trained on the companion dataset: ESO Paranal Real Seeing + Site
Instrument Dataset (2016-2026), 36,706 real hourly observations. See the
dataset card for full column descriptions and provenance.
Training Procedure
- Feature engineering: circular sine/cosine encoding of wind direction;
3-hour rolling means and 3-hour trends (current minus 3-hours-ago) for
temperature, wind speed, vertical-wind turbulence, and pressure; near-
surface temperature gradient; wind shear between 10 m and 30 m.
- Target transform:
log(1 + seeing), to correct a heavy right-skew
(~92% of observations fall between 0.5" and 1.5"; under 0.3% exceed 3").
- Split: chronological, block-based (weekly blocks; whole blocks assigned
to train/validation/test, never split within a block) to avoid leakage
from the strong hour-to-hour autocorrelation of this data. The most recent
15% of weeks were held out as a genuinely unseen future test set.
- Hyperparameter search: Optuna, 80 trials, objective = validation-set
MAE (log scale). Tuned parameters:
n_estimators, learning_rate,
max_depth, subsample, colsample_bytree, min_child_weight,
reg_alpha, reg_lambda, gamma. Early stopping (50 rounds) against the
validation set.
Evaluation Results
| Metric | Value | Notes |
|---|
| R\u00b2 | 0.696 | Chronological hold-out, most recent 15% of weeks |
| MAE | 0.160 arcsec | Real units, after inverse log-transform |
| RMSE | 0.225 arcsec | |
For comparison, a deep neural network trained on the same data/split reached
R\u00b2 = 0.674; linear/ridge/polynomial regression and random forest baselines
all scored substantially lower (R\u00b2 < 0.10 and ~0.45 respectively), consistent
with seeing depending non-linearly on these inputs.
See feature_importance.csv (included) for the full permutation-importance
ranking. The top features are wind direction (sine component), 2 m
temperature, and 30 m wind speed; a 3-hour rolling mean of 10 m wind speed
ranks 4th, confirming recent atmospheric trend carries genuine signal beyond
the instantaneous state.
How to Use
1import numpy as np
2import pandas as pd
3from xgboost import XGBRegressor
4
5FEATURE_ORDER = [
6 "airmass", "pressure",
7 "temp_2m", "temp_30m", "temp_ground",
8 "temp_gradient_ground_30m",
9 "dew_point_depression_2m",
10 "wind_speed_10m", "wind_speed_30m",
11 "wind_shear_10_30",
12 "wind_dir_sin", "wind_dir_cos",
13 "wind_w_std",
14 "rain_intensity",
15 "ir_temperature", "pwv",
16 "temp_2m_roll3h_mean", "temp_2m_trend_3h",
17 "wind_speed_10m_roll3h_mean", "wind_speed_10m_trend_3h",
18 "wind_speed_30m_roll3h_mean", "wind_speed_30m_trend_3h",
19 "pressure_trend_3h", "wind_w_std_roll3h_mean",
20]
21
22model = XGBRegressor()
23model.load_model("xgb_model.json")
24
25# `recent_hours` must be a DataFrame of >= 4 consecutive real hourly
26# observations (oldest first) with the raw columns described in the dataset
27# card, because the rolling/trend features need 3 hours of history - a
28# single isolated snapshot is not sufficient. See inference_example.py for a
29# full worked example including feature reconstruction.
30prediction_log = model.predict(recent_hours[FEATURE_ORDER].iloc[[-1]])
31seeing_arcsec = np.expm1(prediction_log)[0]
See inference_example.py (included) for a complete, runnable example that
reconstructs all engineered features from raw hourly columns.
Citation / Acknowledgement
Trained on data from the European Southern Observatory's public science
archive (http://archive.eso.org). Please credit ESO if you build on this
model or dataset.
