During icing wind tunnel experiments, the calibration process of the spray nozzle and aerothermal systems introduces experimental uncertainty that can potentially compromise the reliability of the test results. Therefore, performing sensitivity analysis (SA) or uncertainty quantification (UQ) studies is not only essential to determine the influence of uncertainties on the ice shape and aerodynamic performance but also crucial to identify the most significant icing parameter uncertainty. However, given the wide range of icing envelopes, it is not practical to conduct SA and UQ by experimental method because a lot of evaluations are required for SA and UQ study. In this study, we addressed these challenges by using a deep learning-based reduced-order modeling technique. First, a dataset covering a wide range of icing envelopes was obtained through icing CFD simulations, and then reduced-order modeling was trained on this dataset to build a model that can efficiently predict the ice shape for the given icing conditions. Finally, the uncertainties in icing conditions were propagated using the ROM model, and SA and UQ were conducted on the various icing conditions covering the icing envelopes of 14 CFR Appendix C to Part 25 and 29. Throughout this paper, we first confirmed that the icing CFD simulation predicts the sensitivity of the experimental results to a reasonable extent in terms of the coefficient of variance. In addition, we confirmed that the prediction error of ROM is within 3% in terms of the ice shape parameters, such as horn thickness and angle. From the SA and UQ study, we found that the effect of uncertainty in ice conditions varies significantly depending on the reference icing conditions. For example, the effect of temperature uncertainty was found to be almost non-existent for low temperature conditions of 245K, but dominant for high temperature conditions of 267K. Finally, the effect of uncertainty in icing conditions was examined for the entire icing envelope.