Improving satellite remote sensing estimates of the global terrestrial water cycle via neural network modeling

Satellite remote sensing provides important observations of Earth’s water cycle, but combining different satellite datasets often fails to produce a balanced water budget, highlighting the errors and uncertainties in these observations. This study introduces a novel approach combining optimal interpolation with neural network modeling to improve global water cycle estimates. We first balance water budget components (precipitation, evapotranspiration, runoff, and water storage change) across 1,358 river basins using optimal interpolation. We then train neural networks to reproduce these results and extend them to ungaged basins. After validating the approach on 340 independent basins, we apply it globally to create calibrated water cycle estimates at 0.5°resolution. Our method significantly reduces water budget imbalances in validation basins, decreasing the mean imbalance from 11 to 0.03 mm/month and reducing its variance from 44 to 24 mm/month. The calibrated datasets perform particularly well when applied to estimating evapotranspiration via the water budget method, achieving accuracy comparable to state-of-the-art methods. This is particularly useful in regions without ground-based measurements, and has broad applications in water resources planning and management. This study helps identify where satellite datasets need correction and demonstrates the benefits of machine learning for studying the water cycle at the global scale. • Machine learning improves the coherency of satellite water measurements globally. • Method cuts errors in water budget calculations compared to existing satellite data. • Technique enables water cycle monitoring in regions lacking ground measurements.