Deep learning based soft-sensor for continuous chlorophyll estimation on decentralized data.

Area of Science:

• Environmental Science
• Data Science
• Sensor Technology

Background:

• Real-time monitoring of water bodies is crucial for conservation.
• Current sensor technology has limitations in real-time, high-frequency measurements for effective risk management.
• Estimating chlorophyll concentration is vital for assessing water quality.

Purpose of the Study:

• To develop a data-driven soft-sensing approach for estimating chlorophyll concentration in river tributaries.
• To evaluate the performance of different neural network learning approaches (individual, centralized, federated) for this estimation task.
• To analyze the impact of data reduction on model performance.

Main Methods:

• Utilized three hydrophysical and three meteorological features as inputs for neural network models.
• Implemented individual, centralized, and federated learning approaches for data-driven estimation.
• Investigated model performance under various data reduction scenarios.

Main Results:

• Individual learning approaches often yielded the best training results.
• Federated learning demonstrated superior generalization ability compared to other methods.
• Federated learning generally improved upon the results obtained with centralized learning.

Conclusions:

• Data-driven soft-sensing offers a viable alternative for estimating chlorophyll concentration when direct measurement is challenging.
• Federated learning is a promising approach for robust water quality monitoring, providing better generalization than centralized methods.
• The findings support the use of advanced machine learning techniques for environmental monitoring and conservation efforts.