An object-based custom ensemble framework for automated water body extraction using Sentinel-2 imagery.

Accurate surface water monitoring across India's diverse landscapes is often hindered by the limitations of pixel-based classification and static indexing. This study introduces an object-based framework integrating DBSCAN clustering with supervised learning classification for automated water extraction using Sentinel-2 imagery. To ensure regional generalizability, the model was trained on a multi-site dataset spanning 20 distinct geographical locations across India. Out of 5 machine learning (ML)models, XGBoost outperformed other ML architectures, including Random Forest and state vector machine (SVM), by effectively minimizing loss in complex spectral zones. Methodological Benchmarking demonstrated that the proposed DBSCAN-XGBoost framework significantly surpassed both traditional NDWI thresholding and pixel-based XGBoost classification, achieving a superior overall accuracy of 97% and an F1-score of 0.96. Feature importance analysis via the Gain metric underscored the synergistic role of mean NDWI and mean NIR in refining decision boundaries. By shifting to an object-oriented paradigm, the proposed model substantially reduced "salt-and-pepper" noise and partially mitigated topographic shadow errors. This research presents a promising object-based framework that, after further multi-site validation, could support high-fidelity water auditing across diverse Indian landscapes.