<![CDATA[This study compares Linear Regression, Random Forest (RF), CNN, and Conv1D-LSTM+Attention for estimating WorldPop-derived population density on 500 m grids in Baghdad and Basra, Iraq (2015–2020). The raw panel contained 13,530 cell-year records, with 12,678 retained after excluding zero-population cells. Alongside conventional random partitions, all four models were evaluated using leave-one-quadrant-out spatial cross-validation. Mean spatial R² was negative for every model in both cities; for example, RF achieved −0.751 ± 1.130 in Basra and −1.096 ± 0.359 in Baghdad. These results contrast with random-split Conv1D-LSTM+Attention performance (R² = 0.833 in Basra; 0.199 in Baghdad), indicating that random partitions overstate out-of-area predictive skill. Moran’s I confirmed strong spatial dependence (Baghdad = 0.847; Basra = 0.946; both p = 0.001). In Basra, a naïve persistence baseline achieved R² = 0.692. Explainability analyses agreed strongly in Basra, where CNN permutation importance and RF-SHAP both identified nighttime lights as dominant; agreement was partial in Baghdad, where CNN ranked LST first while RF-SHAP ranked nighttime lights first. Bidirectional CNN and RF transfer increased RMSE by 33.6–303.8%, indicating poor cross-city portability. Overall, spatial validation is essential for satellite-based population modeling, and WorldPop circularity means the models primarily approximate an existing population surface rather than independent census ground truth.Received: 2026-06-02 Revised: 2026-08-11 Accepted: 2026-08-24 Published: 2026-08-27 ]]>
