Global Digital Elevation Models (GDEMs), such as ASTER, provide near-global coverage at a spatial resolution of 30 m; however, their vertical accuracy varies depending on terrain characteristics, land cover, and sensor-related limitations. This study develops and evaluates a nonlinear Artificial Neural Network (ANN) framework to enhance the vertical accuracy of ASTER DEM within a low-relief urban environment in Baghdad, Iraq. The proposed model is quantitatively compared with deterministic Inverse Distance Weighting (IDW) interpolation. A dataset of 57 high-precision GNSS RTK control points was collected. Forty-seven points were used for ANN training, while 10 independent points were reserved for validation. The ANN model achieved a training RMSE of 4.88 m, while the validation RMSE reached 2.21 m, which is considered the reliable indicator of model performance. Baseline evaluation showed a vertical RMSE of 2.21 m for the original ASTER DEM. After correction, IDW reduced the RMSE to 1.04 m (53% improvement), whereas ANN achieved 0.64 m (approximately 71% improvement), reaching sub-meter vertical accuracy. Residual distribution and comparative statistical analysis confirmed that ANN significantly reduces both systematic bias and error dispersion compared to deterministic interpolation. The findings demonstrate that nonlinear functional modeling substantially enhances the usability of freely available global DEM datasets for high-precision engineering and urban applications. A supplementary cross-validation analysis indicated that this performance advantage is sensitive to the specific data partition used, given the limited sample size; this is discussed as a study limitation, and expanding the training dataset is recommended to enable a more robust comparison.
