Stream salinity prediction in data-scarce regions: Application of transfer learning and uncertainty quantification

Area of Science:

• Environmental Science
• Hydrology
• Data Science

Background:

• Stream salinity data scarcity hinders understanding of water dynamics and management in arid regions.
• Accurate salinity monitoring is crucial for water supply and ecological health.

Purpose of the Study:

• Develop a framework for generating continuous daily stream salinity estimates using instance-based transfer learning (TL).
• Assess the reliability of synthetic salinity data through uncertainty quantification using prediction intervals (PIs).

Main Methods:

• Instance-based TL using Feedforward Neural Networks (FFNNs) calibrated on historical USGS data (1959-1993).
• Testing FFNNs on OWRB data (1998-present) and assessing generalizability in the Bird Creek watershed.
• Utilizing the Lower Upper Bound Estimation (LUBE) method for uncertainty quantification via PIs.

Main Results:

• Autoregressive SC prediction via FFNN showed reliable performance (NSE 0.65 in-sample, 0.45 out-of-sample).
• Bird Creek watershed modeling achieved NSE of 0.54 with similar data scarcity, improving to 0.84 with more data.
• Narrow PIs for North Fork Red River indicated satisfactory salinity predictions (25% range width, 70% confidence).

Conclusions:

• The developed TL framework effectively generates continuous stream salinity estimates in data-scarce environments.
• Uncertainty quantification via PIs demonstrates the reliability of the synthetic salinity data.
• The approach offers a viable solution for water resource management in salt-prone, water-scarce regions.