One-dimensional (1D) site response models assume vertically incident SH waves propagating through laterally uniform soil layers. These assumptions, collectively referred to as the SH1D model, are widely used in site-specific ground motion predictions. However, many studies have demonstrated the limitations of 1D site-response analyses. The term “site response complexity” (SRC) refers to the degree of discrepancy between the observed empirical transfer function (ETF) and the theoretical transfer function (TTF) computed with SH1D modeling. We present a geospatial approach to estimate site response complexity using statistical and machine learning methods with globally or regionally available geospatial proxies. Our site response data are from 114 vertical seismometer arrays in Japan’s Kiban-Kyoshin network (KiK-net) used in Kaklamanos and Bradley (2018). The SRC data are calibrated according to the Thompson et al. (2012) taxonomy that relies on two parameters, r (Pearson’s correlation coefficient between the ETF and TTF) and σi (inter-event variability of the ETF). We examine 18 geospatial proxies associated with site stiffness, topography, basin, and saturation conditions. Using the geospatial proxies as explanatory variables, two sets of predictive models are developed: (a) linear regression models for predicting r and σi, separately, and (b) multiclass classification models for site response complexity. The regression results suggest that predicting σi has greater accuracy than predicting r. Our optimal SRC classification model uses the slope-based VS30 (average shear-wave velocity in the upper 30 m), global sedimentary deposit thickness, and global water table depth as explanatory variables, and has classification accuracies of 0.66 and 0.65 against the training and testing datasets, respectively. We generate maps across Japan for r, σi, and SRC class, separately, which can provide first-order approximations of site response complexity, and exhibit clear patterns between SRC class and topography. We conclude that the geospatial modeling approach is promising for evaluating complexity in site response across broad regions.