A case study of hybrid statistical models for analyzing spatio-temporal effects in historical map series of Donau-Ries county

Abstract

This case study compares generalized linear models, generalized additive models, and semi-structured deep distributional regression for analyzing settlement patterns in historical land-cover maps. Using four reference years for Donau-Ries county (Landkreis Donau-Ries) in southern Germany, we quantify how distances to agricultural areas, forests, and rivers relate to residential presence, and we evaluate predictive performance through spatial block cross-validation. Classical models yield stable and interpretable distance effects, whereas the semi-structured framework improves prediction once local image context is included. The analysis illustrates a practical trade-off between interpretability, accuracy, and robustness when interpretable covariates are combined with deep components in spatio-temporal data analysis.