Predictive Understanding of Stream Salinization in a Developed Watershed Using Machine Learning.

Area of Science:

• Environmental science
• Hydrology
• Machine learning

Background:

• Stream salinization is a growing global concern, impacting aquatic ecosystems.
• Existing models often lack the spatial and temporal resolution for effective ecological exposure assessments.
• Accurate salinity estimates are vital for managing anthropogenic impacts in watersheds.

Purpose of the Study:

• To develop and compare machine learning models for estimating daily stream-specific conductance in unmonitored locations.
• To identify key drivers of stream salinization using explainable artificial intelligence.
• To assess the models' utility in identifying streams impaired by deicer applications for targeted management.

Main Methods:

• Utilized high-frequency monitoring and discrete water quality samples for model training.
• Compared space- and time-unaware Random Forest models with space- and time-aware Recurrent Graph Convolution Neural Network (RGCN) models.
• Applied explainable AI techniques to interpret model predictions and salinization drivers.

Main Results:

• RGCN models achieved comparable predictive performance (KGE: 0.67 and 0.64) to Random Forest models.
• Models accurately captured seasonal salinity patterns, including the winter 'first flush' of deicers.
• High-salinity predictions correlated well with indicators of deicer application and urbanization.

Conclusions:

• Machine learning models can reliably estimate daily stream salinity across watersheds.
• These models effectively identify anthropogenic salinization sources, particularly winter deicer runoff.
• The approach is transferable to other watersheds, supporting targeted best management practices and risk assessment.