Comparison of Shallow and Deep Neural Network Water Temperature Predictions for Resource Management during Cold Stunning Events

In the Laguna Madre, the longest hypersaline lagoon in the United States, the passage of cold fronts can lower air temperature by more than 10°C in less than 24 hours. This can lead to a considerable decrease in water temperature. Records show that some of these cold-water events resulted in massive fish kills and sea turtle cold stunning. In 1997, more than 94,000 fish died in the Lower Laguna Madre and over 48,000 fish died in the Upper Laguna Madre. During the winter of 2010-2011, over 1,600 endangered cold stunned sea turtles were collected while over 3,500 were rescued during the winter of 2017-2018. To mitigate the impact of these cold events, local agencies, private sector companies and other stakeholders interrupt activities such as fishing and navigation during the events. However, to manage interruptions of commercial navigation and preparation of resources ahead of these events accurate predictions are necessary to allow for rerouting cargo and to mobilize volunteers and response assets.

It was found that water temperature is the key environmental variable correlated with fish kills and sea turtle cold stunning, hence the challenge to accurately predict future water temperatures with long lead times. Artificial neural network models were trained and tested over past hourly measurements of water and air temperatures in the Laguna and the nearby coastline of the much larger Gulf of Mexico. The optimization process, the predictors and the neural network topologies selected for the different lead times will be described. The performance was assessed for varying prediction times (3, 12, 24, 48 and up to 84 hours). To assess model performance more specifically during cold events, performance metrics were also assessed during times for target water temperature below 12°C. Results confirm the applicability of the models for cold events. While shallow neural networks are implemented operationally, additional research is ongoing to compare performance of shallow neural networks with deep neural networks applied to the same data. Given the large number of inputs selected for the operational models (over 100 predictors) and the large data sets (over 3 million data points per time series) a deep learning approach was deemed promising. The deep learning technique selected is the stacked denoising autoencoder (SDAE) whereby hidden layers are pretrained in an unsupervised manner in order to create a higher order representation of the features. The optimal SDAE architecture will be described along with a comparison of the results of the two types machine learning approaches, shallow and deep neural networks.