Optimized Wetland Classification in Arid Coastal Environments: Integrating Sentinel-2 Imagery with Hyperparameter-Tuned Machine Learning Algorithms

Abstract. Coastal wetlands play a vital role in maintaining ecological balance by providing essential ecosystem services such as carbon sequestration, biodiversity conservation, and water regulation. However, these sensitive environments are increasingly threatened by human pressures and climate change, underscoring the need for accurate, scalable, and efficient monitoring approaches. Remote sensing offers cost-effective alternatives to traditional wetland monitoring methods. In this study, we developed an advanced machine learning framework for classifying and analyzing the Miankaleh coastal wetland an internationally recognized Ramsar site located along the southeastern coast of the Caspian Sea using Sentinel-2 data processed on the Google Earth Engine (GEE) platform. In this context, four machine learning algorithms Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Ensemble Learning—were systematically evaluated for wetland classification after extracting diverse spectral, spatial, and textural features, including vegetation and water indices. The findings revealed that among the tested classifiers, XGBoost achieved the highest accuracy (OA = 0.86, Kappa = 0.835) with the shortest computation time (49 seconds), outperforming traditional methods. Furthermore, hyperparameter optimization using Grid Search, Random Search, Bayesian Optimization, and SHAP-based tuning showed that although Grid Search produced the highest Kappa (0.839), its computational cost was more than eight times greater than the default configuration. These results demonstrate that XGBoost, even with minimal tuning, provides an optimal balance between classification accuracy and computational efficiency for coastal wetland environments. The proposed framework highlights the potential of integrating open-access satellite data, cloud-based processing, and optimized machine learning models for large-scale, operational wetland monitoring and management.