Assessing the Influence of Sample Size and Spatial Sampling Strategies on GeoAI-Based Digital Elevation Model Error Correction: A Case Study of the Sabaragamuwa University Area, Sri Lanka

Digital Elevation Models (DEMs) are essential for terrain analysis, hydrological modeling, disaster assessment, and environmental monitoring. However, freely available DEMs often contain significant vertical errors due to the terrain complexity, vegetation cover, and sensor-related limitations. Although Machine Learning (ML) techniques have strong potential for DEM corrections, the effect of sample sizes and sampling strategies under limited ground data
conditions remains insufficiently studied. Therefore, this study investigated the influence of sample sizes and sampling strategies on Geo-Artificial Intelligence (GeoAI)-based DEM error correction within the Sabaragamuwa University area, Sri Lanka. A Random Forest (RF) regression model was developed by using the elevation, slope, aspect, and roughness derived from the Shuttle Radar Topography Mission (SRTM) DEM as predictor variables, while DEM vertical error was set as the target variable. Ground survey elevations collected through Total Station and Automatic Level instruments were used as reference data. To simulate the practical limitations in field elevation data acquisition, multiple sample sizes, ranging from 50 to 500 points, were evaluated under random, clustered, and stratified sampling strategies. Model performance was evaluated using Root Mean Square Error (RMSE), Coefficient of Determination (R²), percentage improvement, and variogram-based geostatistical analysis. The results demonstrated a significant improvement in DEM accuracy after correction. The original DEM RMSE of 5.010m was reduced to values between 1.973m and 3.186m. The stratified sampling strategy consistently produced the best overall performance (RMSE =1.973m, R²= 0.575). Variogram analysis further confirmed a reduction in
spatially autocorrelated errors in the corrected DEM, while independent validation using reserved ground survey points further verified the reliability of the proposed framework. Overall, the findings demonstrated the effectiveness of the proposed GeoAI-based approach for cost-efficient DEM correction with limited ground data.