Forecasting Cloud Cover and Atmospheric Seeing for Astronomical Observing: Application and Evaluation of the Global Forecast System

To explore the issue of performing a non-interactive numerical weather forecast with an operational global model in assist of astronomical observing, we use the Xu-Randall cloud scheme and the Trinquet-Vernin AXP seeing model with the global numerical output from the Global Forecast System to generate 3-72h forecasts for cloud coverage and atmospheric seeing, and compare them with sequence observations from 9 sites from different regions of the world with different climatic background in the period of January 2008 to December 2009. The evaluation shows that the proportion of prefect forecast of cloud cover forecast varies from ~50% to ~85%. The probability of cloud detection is estimated to be around ~30% to ~90%, while the false alarm rate is generally moderate and is much lower than the probability of detection in most cases. The seeing forecast has a moderate mean difference (absolute mean difference <0.3" in most cases) and root-mean-square-error or RMSE (0.2"-0.4" in most cases) comparing with the observation. The probability of forecast with <30% error varies between 40% to 60% for entire atmosphere forecast and 40% to 50% for free atmosphere forecast for almost all sites, which being placed in the better cluster among major seeing models. However, the forecast errors are quite large for a few particular sites. Further analysis suggests that the error might primarily be caused by the poor capability of GFS/AXP model to simulate the effect of turbulence near ground and on sub-kilometer scale. In all, although the quality of the GFS model forecast may not be comparable with the human-participated forecast at this moment, our study has illustrated its suitability for basic observing reference, and has proposed its potential to gain better performance with additional efforts on model refinement.

💡 Research Summary

The paper investigates whether a fully automated, non‑interactive weather forecast can be used to support astronomical observing by exploiting the operational Global Forecast System (GFS). The authors couple two physics‑based schemes to the GFS output: the Xu‑Randall cloud parameterisation for predicting cloud cover, and the Trinquet‑Vernin AXP model for estimating atmospheric seeing (optical turbulence). Forecasts are generated for lead times of 3, 6, 12, 24, 48 and 72 hours and are compared with two‑year (January 2008–December 2009) observational records from nine sites distributed across diverse climatic regimes (e.g., African savanna, South‑American Andes, East‑Asian plateau, European coastal stations).

For cloud cover, the proportion of “perfect forecasts” (exact match between forecast and observation) ranges from roughly 50 % to 85 % depending on the site. The probability of detection (POD) varies between ~30 % and ~90 %, indicating that the model is generally capable of identifying cloudy conditions when they occur. The false‑alarm rate (FAR) is consistently lower than POD, meaning that missed clouds are more common than spurious cloud predictions. These statistics reflect the GFS’s strength in capturing large‑scale synoptic patterns but also its limitation in resolving thin, low‑level cloud layers that are critical for astronomical scheduling.

Seeing forecasts show a moderate level of skill. The absolute mean difference between forecast and measured seeing is below 0.3 arcseconds for most sites, and the root‑mean‑square error (RMSE) lies between 0.2″ and 0.4″. When the forecast error is expressed as a percentage of the observed value, the probability of achieving <30 % error is 40 %–60 % for total‑atmosphere seeing and 40 %–50 % for free‑atmosphere seeing. This performance places the GFS/AXP combination in the better cluster among existing seeing models, especially for free‑atmosphere turbulence where ground‑layer effects are excluded.

Nevertheless, a subset of sites—particularly high‑altitude or desert locations—exhibit considerably larger errors. The authors attribute these outliers primarily to two model deficiencies: (1) the coarse horizontal (≈28 km) and vertical resolution of the GFS, which cannot resolve sub‑kilometer scale temperature and moisture gradients that drive low‑level turbulence; and (2) the AXP model’s reliance on bulk atmospheric stability and wind profiles without explicit representation of local terrain, surface roughness, or land‑surface processes. Consequently, the model underestimates turbulence near the ground and on scales smaller than the GFS grid, leading to systematic seeing biases at those stations.

The study concludes that, while the GFS‑based forecasts do not yet match the accuracy of human‑in‑the‑loop, expert meteorological forecasts, they are sufficiently reliable for basic observing reference—providing a quick, objective assessment of cloud probability and seeing quality. The authors suggest several pathways for improvement: (i) nesting higher‑resolution regional models (e.g., WRF) within the GFS framework to capture mesoscale phenomena; (ii) augmenting the AXP turbulence parameterisation with site‑specific terrain and surface‑property inputs; and (iii) assimilating real‑time observations such as satellite cloud indices, ground‑based lidar, or all‑sky cameras to refine initial conditions. With these enhancements, an operational, fully automated forecast system could become a valuable tool for observatory scheduling, instrument protection, and remote‑site decision making.