A Cloud-Native Google Earth Engine Framework for Water Accounting Land Use (WALU) Classification: Application to the Narmada River Basin

Sustainable management of river basins requires land and water to be understood as a single interacting system, yet the land-use datasets most widely used in basin planning primarily represent physical land cover and rarely distinguish hydrologically distinct land uses, grouping hydrologically distinct pixels (e.g., rainfed and irrigated cropland, or natural and engineered water bodies) into a single undifferentiated class. The Water Accounting Land Use (WALU) classification of the Water Accounting Plus (WA+) framework was designed to close this gap, but its conventional implementation has typically been implemented as a desktop workflow that loads complete rasters into local memory, which limits its efficiency for large basins and fine-resolution inputs. This study addresses that constraint by re-engineering the WA+ WALU workflow as a cloud-native, server-side classification on Google Earth Engine, integrating nine geospatial input layers through a hierarchical rule-based fusion that implements the original WA+ classification rules, each checked against the desktop implementation, while avoiding the need to load complete raster datasets into local memory and providing a scripted workflow that can be re-run by other users. The 24-class, pixel-level WALU map of the 97,366 km2 Narmada River Basin reveals that the human-modified land uses Modified Land Use (MLU) and Managed Water Use (MWU) (51.5%) marginally exceed the natural and semi-natural land uses Protected Land Use (PLU) and Utilized Land Use (ULU) (48.5%). Irrigated cereals (MWU3; 32.21%) form the largest WALU class, exceeding closed deciduous forest (23.41%), which indicates a strong dependence of agricultural production on irrigation and increased engineered water regime. The classification also maps wasteland over 9.91% of the basin, providing a spatially explicit basis for prioritizing land-restoration assessments, and shows marked sub-basin and district-scale differences in land use, with irrigated cropland concentrated in about ten districts within the major reservoir command areas against a periphery of rainfed and degraded land. A two-level accuracy assessment, combining a stratified confusion-matrix analysis (86.4% area-weighted accuracy; Kappa = 0.82) with a Sentinel-2 NDVI (Normalized Difference Vegetation Index) phenological validation of the rainfed-irrigated split (Cohen’s d = 1.27), indicates that the map is suitable as a land-use input for basin-scale WA+ accounting at the aggregated class level. The framework can be adapted for WALU mapping across other data-scarce basins.