Delineating small water bodies in Pune City India using Machine Leaning in Google Earth Engine

Freshwater is a valuable and scarce resource under constant threat due to global climate uncertainties, population growth, and economic expansion. Mapping water bodies can be useful in effective water resources management. The present study was an effort towards mapping inland water body and identifying areas suitable for water conservation in the Pune city of the Upper Bhima River Basin. In this study, land cover classification was performed using machine learning in Google Earth Engine (GEE) to identify water and non-water pixels to delineate small water bodies. Three machine learning models, namely Support Vector Machine (SVM), Random Forest (RF) and Gradient Tree Boost (GTB), were compared for their efficacy in mapping the water bodies. An open-source high-resolution multi-spectral image (MSI) information from Copernicus Sentinel-2 Level 2A harmonised data was used to generate a water body map. The classification models were further compared with the Modified Normalized Difference Water Index (MNDWI) thresholding method, which distinguishes water regions based on the reflectance difference between the Short Wave Infra-Red (SWIR) band and the Green band. As the study area covered a diversified spectral signature of land use and land cover, the analysis was performed under three scenarios. In scenario 1, the ML model was trained and validated using hilly and built-up region data, in scenario 2 agricultural and built-up areas were considered and in scenario 3 all three regions were covered. Results showed that the SVM model performed more accurately and detected the maximum area of water bodies followed by RF, GTB and MNDWI threshold methods. Moreover, scenario 3 which considers the entire dataset ranging from hilly, built-up and agricultural regions is the most robust analysis to perform water body mapping. Finally, the SVM model considering scenario 3, was used to detect the small water bodies for the entire catchment. In total, 20,479 water bodies were identified by the SVM model covering 279.42 sq.km area. Furthermore, river networks were removed from the classification, which resulted in a total of 17,616 small water bodies with an area of 243.97 sq.km. As this analysis was performed using Sentinel-2A data which has spatial resolution of 10 meters, ML models and MNDWI method cannot estimate water bodies smaller than 100 sq. meters. The water body map can be useful for water resources planning in the study area.

Keywords: Google Earth Engine, Random Forest, Support Vector Machine, Gradient Tree Boost and Modified Normalized Difference Water Index.