Uncertainty Quantifications of the Onset and Offset of Cold-Stunning Events Using AI Ensemble Methods

The Laguna Madre, a large estuarine system located off the coast of southern Texas, is home to numerous protected species including the endangered green sea turtle. However, given the weather dynamic of southern Texas, the Laguna Madre can cool down very rapidly and serve as a “trap” to marine life when cold weather fronts travel to the coast. This results in the loss of mobile capabilities for economically valuable fish and endangered sea turtles, potentially leading to severe cold-related illnesses or death. From 1980 to 2015, ~8100 stranded sea turtles were recorded in Texas, and approximately ~56% of stranded turtles found had undergone hypothermic stunning or “cold-stunning” (CS). This places heavy emphasis on effectively predicting and communicating when water temperatures fall below CS thresholds to provide stakeholders with sufficient time to prepare for sea turtle and fishery recovery efforts and mitigation measures such as navigation interruptions in the Laguna Madre.

Conrad Blucher Institute (CBI) AI researchers at Texas A&M University-Corpus Christi (TAMU-CC) have developed an existing operational model that predicts water temperatures in the Laguna Madre. These predictions have allowed coastal stakeholders along the Texas coast to try to minimize the impact of these rare events by organizing volunteers and partners for turtle recovery efforts and by interrupting navigation and/or engineering work when water temperatures drop below the critical CS threshold. However, stakeholders have communicated a need for an improved model that provides uncertainty estimates of these water temperature predictions to assist with precautionary multi-domain decision-making processes before these rare cold water events occur.

In order to more accurately communicate when and how long water temperatures will fall below the critical cold-stunning threshold (≈8˚C for sea turtles and ≈4.5 ˚C for fish; Shaver et al., 2018), the project focuses on estimating onset and offset CS threshold-crossing variabilities as well as newly constructed confidence intervals based on an ensemble of neural networks (NNs). The NNs are trained based on ten years of air and water temperature data (TCOON) and assessed using a k-fold cross-validation method based on yearly stratifications. The confidence intervals around the timing of threshold-crossings will be created by extracting the earliest and latest crossings of the ensemble members. Ensemble members will be constructed by (1) repeating calibration of the NN models and (2) perturbing the air temperature predictions used as inputs for model calibration to simulate a range of warmer and cooler air temperature predictions during such events for the same location.

Creating a new model with better prediction accuracy and explainable capacity will not only help preserve marine life in south Texas waters and possibly other locations, but also further improve the communication framework and decision-making processes with key stakeholders (i.e., National Park Service, Texas Parks and Wildlife, Coastal Conservation Association, Gulf Intracanal Association, etc.).