Estimating River Water Quality Indicators in South Korea Using AlphaEarth Embeddings

Water quality monitoring is essential for river management and ecosystem protection, yet in situ networks remain sparse relative to the spatial and temporal variability of rivers.This study evaluates whether geospatial foundation model embeddings can support scalable river water quality estimation.We combined annual observations from South Korea's national monitoring network (2017-2024) with AlphaEarth Foundation embeddings (64-dimensional latent features at 10 m resolution) and trained gradient-boosted trees to predict eight indicators.Reduced feature sets matched or exceeded the full model, while excluding spatial coordinates degraded performance, underscoring their importance.Across indicators, models achieved robust agreement with in situ data (overall out-of-sample R0.62-0.89).Pixelwise predictions resolved coherent spatial structure and interannual variation across four major rivers, though narrow shoreline edges displayed atypical values.Recurring embedding features were identified across multiple targets, suggesting that AlphaEarth encodes latent properties of inland waters.While limited by the available temporal span and spatial coverage, the results demonstrate that geospatial foundation embeddings enable accurate and transferable estimation of river water quality and provide reusable features for future aquatic and environmental applications.