A Multi-District GeoAI Framework for Forecasting Spectral Built-up Density Index Change: Predictability Limits and Uncertainty in Baghdad, Iraq

Short satellite archives may contain limited information for forecasting urban surface change. This study benchmarks a Spectral Built-up Density Index (SBDI) across three spatially disjoint Baghdad districts using 8,015 cells of 100 m × 100 m observed biennially during 2014–2024. Nine predictors, including persistence baselines, Random Forest (RF), LSTM, gradient boosting, Gaussian process regression and an RF–LSTM ensemble, are evaluated in a 2024 hindcast with block spatial cross-validation. The best models achieve R² = 0.505, 0.738 and 0.778 in Al-Jadriya, Al-Karada and Al-Adhamiya, respectively; persistence with a training-fold drift constant achieves 0.490, 0.720 and 0.742. The block-bootstrap interval for RF-engineered improvement over persistence with drift excludes zero only in Al-Adhamiya (mean ΔR² = +0.024; 95% interval +0.003 to +0.044). R² on change remains at or below 0.131. Leave-one-district-out RF transfer yields R² = 0.508–0.753. Separate spatial training, calibration and test folds yield 89.2–90.3% empirical coverage for nominal 90% hindcast intervals. Applying the pooled hindcast calibration to exploratory 2026 forecasts identifies 50 cells whose intervals exclude their observed 2024 values, without a temporal coverage guarantee. Most Local Moran's I clusters disappear after false-discovery-rate control, and hindcast cluster agreement is weak. The framework supports assessment of predictability and targeted verification of spectral change; it does not establish deterministic forecasts of physical urban intensification.