Remote sensing-enabled machine learning for river water quality modeling under multidimensional uncertainty.

Area of Science:

• Environmental Science
• Hydrology
• Remote Sensing
• Machine Learning

Background:

• River water quality simulation is hindered by insufficient data and models that cannot capture complex influencing factors.
• Traditional Total Suspended Solids (TSS) data collection is sparse, limiting analysis of climate extremes and water contamination.
• Existing models struggle with the multidimensionality of factors affecting river water quality, including hydro-climatic conditions and land use.

Purpose of the Study:

• To develop a technique augmenting limited ground-based water quality observations with remote-sensed data.
• To leverage machine learning to account for complex, multidimensional influences on river water quality.
• To improve the prediction of water quality variables, such as Total Suspended Solids (TSS), under climate uncertainty.

Main Methods:

• Utilized a Long Short-Term Memory Network (LSTM) model.
• Trained the LSTM model on Moderate Resolution Imaging Spectroradiometer (MODIS) satellite reflectance data.
• Calibrated the model using Total Suspended Solids (TSS) data from the Ohio River Valley Water Sanitation Commission (ORSANCO).

Main Results:

• Successfully augmented limited ground-based water quality data with remote-sensed reflectance data.
• Developed a data-driven algorithm capable of accounting for spatial variability within a watershed.
• Demonstrated effective water quality prediction under uncertainty, using TSS as a surrogate for pollutants.

Conclusions:

• The proposed methodology enhances water quality simulation by integrating satellite data and machine learning.
• This approach addresses data scarcity and model complexity issues in environmental monitoring.
• The developed technique provides a robust framework for empirical analysis and data-driven water quality prediction.