A case study of hybrid statistical models for analyzing spatio-temporal effects in historical map series of Donau-Ries county
In the authors' words
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.
Appeared: Friday, October 30. arXiv. Preprint, not yet peer-reviewed.
Authors' comment: 28 pages, 8 figures. Preprint prepared for Communications in Statistics: Case Studies, Data Analysis and Applications