Using Continuous Rank Probability Score Multi-Model Ensemble Machine Learning Method to Quantify Aleatoric and Epistemic Uncertainty for Improvement of Cold-Stunning Predictions

In the Laguna Madre, an estuary in southern Texas, major cold-stunning events occur when water temperatures fall below 8°C for about 24 hours, leaving endangered green sea turtles lethargic and potentially dying if not rescued. During the winter of 2020-2021, a record 13,000+ sea turtles were affected by a severe cold-stunning event. To help minimize the impacts of these events, an operational machine learning (ML)-based model has been developed by the Conrad Blucher Institute (CBI) that outputs a single deterministic water temperature 120-hour lead time water temperature prediction. To enhance the utility of these predictions for stakeholders, other ML models have also been developed to quantify the uncertainty of the ML-derived water temperature predictions, including a Multi-Model Ensemble (MME) that estimates epistemic (model) uncertainty and a Probabilistic Neural Network (PNN) that estimates aleatoric (data) uncertainty. The aim of this research is to apply a new ML Uncertainty Quantification (UQ) method that implements the Continuous Rank Probability Score (CRPS) as the loss function within an MME model to gain an estimate of both aleatoric and epistemic uncertainty of the water temperature predictions simultaneously. By using UQ methods, more information can be gathered to gain better insight into helping improve mitigation strategies for cold-stunning events. Future work will involve a comparative analysis of all three UQ methods (i.e., MME, PNN, and CRPS-MME) to assess the uncertainty estimates for cold-stunning events.