Enhancing environmental data imputation: A physically-constrained machine learning framework.

Area of Science:

• Environmental Science
• Water Resources Management
• Machine Learning

Background:

• Integrated watershed models are crucial for analyzing climate and water resources.
• Limited field data introduces significant uncertainty into these models.
• Exploiting available data is key to improving model performance before collecting more.

Purpose of the Study:

• To develop a novel machine learning framework for imputing missing environmental data.
• To assess the framework's effectiveness in handling data gaps across various domains.
• To enhance the performance of integrated environmental models through data augmentation.

Main Methods:

• A machine learning framework incorporating physical constraints was developed.
• The framework was applied to impute missing data in meteorology, water quantity, and water quality.
• Model performance was evaluated using the Nash-Sutcliffe Efficiency (NSE) metric.

Main Results:

• The framework successfully imputed a high percentage of missing data across environmental domains.
• Satisfactory imputation results were achieved, with minimum NSE values of 0.72 for meteorology and >0.97 for hydrometric variables.
• Over 78% of physical-water-quality variables and 66% of chemical-water quality variables showed NSE > 0.45 and > 0.35, respectively.

Conclusions:

• The proposed machine learning framework is effective for data augmentation in environmental modeling.
• Imputing missing data significantly improves the performance and reduces uncertainty in integrated watershed models.
• This approach offers a valuable tool for optimizing water resources management through enhanced data availability.