Lowering barriers to probing high-frequency variations in river chemistry through a frugal machine learning-based framework

Amita Prajna Mallik , Antoine Lucas, Eric Gayer, Jérôme Gaillardet

High-frequency river chemistry monitoring is crucial for capturing transient hydro-geochemical variations and ensuring water security, yet its implementation is limited by logistical and budgetary constraints. Here we present a machine learning-based framework that integrates continuous, low-cost physico-chemical proxies with sparse ‘anchoring’ solute measurements to reconstruct hourly-scale variations in major dissolved metals and nutrients. Validated across three contrasting catchments, we demonstrate that daily to semi-weekly sampling suffice to achieve accurate reconstructions (NSE > 0.75), though sampling demand varies considerably across hydrological regimes. Prediction of nutrients (e.g. NO₃⁻ and K⁺) may require more frequent anchoring observations under stormflow conditions. Reconstruction accuracy is sustained even under multi-week voids in anchoring observations, demonstrating resilience to commonplace logistical disruptions. We complement this framework with a novel gap imputation method based on Singular Spectrum Analysis to address missing data in proxy time series, which outperforms traditional gap-filling approaches. Our results support a scalable, low-maintenance strategy that enables a >95% reduction in operational costs and carbon emissions associated with high-frequency monitoring; and provide a transferable, hydrological regime-specific roadmap for optimizing field sampling that minimizes logistical burden while maintaining reconstruction accuracy.