Enhancing GeoAI land cover classification via hyperparameter tuning and cross-validation: A case study in Ravenna, Italy

Sustainable land management requires accurate and timely monitoring of Land Cover (LC), especially in rapidly evolving urban areas. Sentinel-2 imagery provides valuable multispectral data for LC classification; however, distinguishing spectrally similar LC types remains challenging. This study proposes a methodological framework for enhancing LC classification accuracy by utilizing Artificial Intelligence for Geospatial Data (GeoAI). It evaluates the performance of Random Forest (RF) and Support Vector Machine (SVM) classifiers, optimized through hyperparameter tuning and cross-validation. The classification was implemented on Google Earth Engine (GEE) using 2024 Sentinel-2 imagery for Ravenna, Italy (630 km2), distinguishing six LC classes: water, tree cover, grassland and shrubland, cropland, built-up areas, and bareland. Four experimental scenarios were examined: baseline, cross-validation (CV), hyperparameter tuning (Tuning), and their combination (CV + Tuning). The Classification accuracy was assessed using several metrics, including Overall Accuracy (OA), Kappa Coefficient (KC), Producer's Accuracy, User's Accuracy, and F1-Score. The results were statistically validated using McNemar's test, a novel contribution that ensured the statistical effectiveness of the observed accuracy improvements. Results demonstrate RF consistently outperformed SVM across all scenarios. In the fully enhanced scenario (CV + Tuning), RF achieved an OA of 92.5 % and KC of 0.91, approximately 10 % higher than SVM (82.0 % OA, KC 0.78). McNemar's test confirmed statistically significant accuracy differences for most LC classes. These findings highlight that systematically integrating hyperparameter tuning and cross-validation significantly enhances classification reliability, particularly using RF. Future research should explore additional classifiers and incorporate ancillary data sources to further improve the methodological robustness of LC mapping.