Using 23 Years of ACS/SBC Data to Understand Backgrounds:Explaining & Predicting Background Variations

TL;DR

By analyzing 8,640 SBC exposures with 23 observational parameters, the study builds a robust Quantile Random Forest model to predict SBC backgrounds and uses SHAP explanations to identify key drivers. The results show that Solar position and SBC temperature dominate background variations in different filters, with geocoronal effects generally modest and Galactic latitude contributing to higher backgrounds near the plane. The work provides actionable guidance for planning ACS/SBC observations to minimize background, such as favoring night-time conditions, enforcing minimum Earth limb angles, and accounting for instrument temperature in scheduling. Overall, the paper delivers a principled framework for predicting and interpreting SBC backgrounds, improving exposure planning and data quality for UV observations.

Abstract

Recent analysis of 23 years of Hubble Space Telescope ACS/SBC data has shown that background levels can vary considerably between observations, with most filters showing over an order of magnitude variation. For the shorter-wavelength filters, the background is understood to be dominated by airglow; however, what precisely drives background variations is not well constrained for any filter. Here, we explore the causes of the background variation. Using over 8,000 archival SBC observations, we developed a machine learning model that can accurately predict the background for an observation, based upon a set of 23 observational parameters. This model indicates that, depending on filter, the SBC background is generally dominated by Solar elevation, Solar separation angle, Earth limb angle of observation, SBC temperature, and target Galactic latitude.

Figures (21)

• Figure 1: Map showing the position of HST during every SBC exposure considered in this study. Each exposure is plotted with 3 points, corresponding to its start, mid-point, and end. All points are color-coded according to the strength of the geomagnetic field at that position, as per IGRF-14. Orbits that would intersect the SAA, and where no SBC exposures were therefore taken, are neatly illustrated by the gaps in this map. The region around the SAA still has a notably weaker geomagnetic field than elsewhere.
• Figure 2: Comparison of predicted versus true background levels, for each SBC filter, from our quantile forest regression modelling. For each filter's data, we trained using 80% of the observations, and withheld the remaining 20% for validation testing. For each filter, the training data is plotted in orange, and the test data is plotted in blue (along with the uncertainties on the predictions, as provided by the model). Prism data is shown only for completeness.
• Figure 3: SHAP dependence plots for Solar altitude for a representative subset of SBC filters. The left-hand y-axis shows the marginal impact of the parameter on the model prediction in ${\rm counts\,sec^{-1}\,arcsec^{-2}}$, whilst the right-hand y-axis shows the impact on the model normalized to the median background for that filter (the median background for the filter in question is noted on each plot). The y-axes are plotted using a symlog scale (linear either side of 0, and logarithmic elsewhere). The shaded region shows the 90th percentile range in SHAP dependence, averaged within a moving Gaussian window with sigma equal to 5% of the parameter range. Shown for comparison is a plot of measured background versus Solar altitude for F125LP observations, to illustrate a noisier raw relation, as compared to the marginal impact derived by SHAP analysis.
• Figure 4: SHAP dependence plots for several relations where our QRF regression model has not constrained how a parameter impacts the background, as indicated by large scatter, either in localized spikes and jumps, or throughout larger portions of the parameter range. Details otherwise as per Figure \ref{['Fig:Solar_Alt_Dependence_Plots']}.
• Figure 5: Beeswarm plot for SHAP analysis of our QRF regression model for F125LP and F150LP. The rows are ordered according to the tightness of the correlation between the parameter in question, and the impact on the model (calculating using the Kendall's tau correlation coefficient; Kendall1990). This means that parameters located lower in the plot are more likely to have poorly-constrained or spurious impacts on the model predictions.