Constructing a Two-Dimensional Ultraviolet Atmospheric Transmittance Field Using Gaussian Process Regression

Abstract. Accurate characterization of slant-path atmospheric transmittance in the 280–400 nm ultraviolet (UV) window is essential for radiometric calibration of ground-based UV observations and link-budget assessments of UV optoelectronic systems. Nighttime measurements, however, typically rely on stellar radiometers that provide only sparse and irregular samples along stellar tracks, which complicates construction of a continuous two-dimensional (zenith–azimuth) transmittance field. Here we retrieve slant-path transmittance from multi-star, multi-channel stellar UV radiometer observations using a stellar Langley-type calibration, and reconstruct a 2-D transmittance field by modeling optical depth with Gaussian process regression (GPR) as a function of zenith angle and azimuth. Predictive uncertainty is quantified and rescaled using cross-validated standardized residuals, and performance is benchmarked against conventional interpolation approaches. The reconstructed fields robustly capture the zenith-angle-dominated gradient, achieving cross-validated R² = 0.965–0.984 and RMSE = 0.004–0.007 across four UV channels. After calibration, 67 %–81 % of standardized residuals fall within ±1 and 93 %–96 % within ±2 standard deviations, indicating well-calibrated uncertainty; uncertainty increases in sparsely sampled regions and with local sampling spacing. The proposed framework enables practical construction of nighttime UV transmittance fields with well-characterized uncertainty for observation correction and quantitative assessment of site transparency.