The national seismic networks of Switzerland comprise more than 200 stations. At the station sites, the empirical amplification functions (EAFs) are routinely computed after each earthquake using a generalized inversion method based on separation of source, path, and site effects. The seismic stations are also characterized through geophysical measurements aiming to estimate shear‐wave velocity profiles and horizontal‐to‐vertical spectral ratio of ambient vibrations (HVNR). Using this information, the correlation between the HVNR and EAF is assessed through canonical correlation. Once established, the canonical correlation is used to reconstruct the expected at each considered station site in the dataset. The prediction is individually made for all seismic stations in the dataset, excluding every time the investigated station is from the calibration dataset; the reconstruction of the is performed resorting to two parallel methods. The first method uses a combination of the canonical correlation parameters and Moran index, and the second one solves in a least‐squares sense an overdetermined linear equation system including the canonical couples deemed as reliable. After a first round of predictions, a systematic lower in soft sediment sites and a higher in hard‐rock sites is observed. A possible explanation for this behavior is found in the “normalization” to the Swiss standard rock profile in the computation of the EAF at the Swiss stations. Therefore, to reduce this effect, geological and geophysical parameters are considered in addition to the HVNR in the canonical correlation. We observe that the final solution improves when the least‐squares solution approach is used with a combination of HVNR, , and thickness of the ice cover at the last glacial maximum. Moreover, a blind test is performed using data not considered in the calibration dataset. The results highlight the ability of the method to provide an estimate of the site amplification over chosen frequency bins.