Analysis on Machine Learning Applications and Challenges in Surface Water Quality Management

Abstract Submission: Surface waterbodies are heavily exposed to pollutants caused by natural disasters and human activities. Empowering advanced sensor technologies in water quality monitoring, sufficient measurements are available to develop machine learning (ML) models for surface water management. Due to the rapid ML development, it is challenging to choose the appropriate algorithms and strategies for ML water quality models. After synthesizing recent studies, we categorized ML water quality models into three main groups: point-to-point (P2P), sequential-to-point (S2P), and sequential-to-sequential (S2S). P2P models estimate current target values based on other measurements at the same time point. They were commonly built by Random Forest (RF) and shallow Artificial Neural Network (ANN) and performed well in many cases for Water Quality Index/Classification (WQI/WQC), Dissolved Oxygen (DO), Conductivity, Phosphorus, etc. S2P models utilize previous time series data to predict the target value at only one-time point ahead. Long Short-Term Memory (LSTM) was widely used and obtained favorable results for developing S2P models. S2S models forecast the following time period target values depending on the previous time series data, providing a detailed changing trend. Advanced deep learning (DL) models such as attention-based Gate Recurrent Unit (AT-GRU), AT-LSTM and Transformer were generally selected for S2S models and achieved acceptable performance. In conclusion, recommendations on ML model selections highly depends on the data availability, model performance and water management needs.

Learning Objectives/Expected Outcome (Optional) :