All models used the same three explanatory variables – distances to agriculture, forest and river – derived from historical land‑cover maps for four reference years.

The GLM treats each distance as a linear predictor and estimates odds ratios for being settled. The GAM adds smooth functions for each distance and a two‑dimensional smooth for spatial coordinates, allowing more flexible spatial patterns. The SDDR embeds neural network layers within the GAM framework, adding an image‑based neighbourhood that can capture local context.

To evaluate the models, researchers partitioned the county into four spatial blocks and performed block cross‑validation. In that test, the GLM achieved an accuracy of 0.7556, precision 0.3784, recall 0.9774 and F1‑score 0.5200. The GAM performed worse in cross‑validation (accuracy 0.6979, precision 0.3543, recall 0.6240, F1‑score 0.3327) despite a high in‑sample R² of 0.93, suggesting over‑fitting to spatial patterns.

The most powerful SDDR variant – Model 3 with a 17×17 pixel neighbourhood – reached an accuracy of 0.9490, precision 0.7940, recall 0.7125 and F1‑score 0.7501. Smaller neighbourhoods or models lacking image context (SDDR 1 and 2) performed poorly, with accuracies only 0.44–0.48.

The analysis is limited to the three distance covariates and does not account for socioeconomic drivers such as population growth or policy changes. Therefore, the reported associations are conditional and not causal.

These results show that incorporating local image context into distributional regression yields substantially better settlement predictions than traditional statistical models in this setting.