4:45 PM - 5:00 PM
[ATT35-12] Detection of Irminger Rings in high resolution ocean hydrodynamic modeling data using artificial neural networks
Keywords:Subpolar North Atlantic, eddy-resolving modelling, eddy identification, artificial neural networks, Bayesian optimization
Deep convection in the Labrador Sea is a key component in the formation of the lower branch of the Atlantic Meridional Overturning Circulation (AMOC). It is known that mesoscale eddy activity in the Labrador Sea, represented by Irminger Rings (IR), influences the convection process. In order to analyze the impact of IRs on the spatial-temporal variability of the mixed layer depth, it is necessary to create a trajectory database of eddy motion, which poses the problem of IRs detection and tracking with high accuracy. In this study, we propose the novel technique for detection of IRs in high-resolution ocean numerical simulation. The research is based on the regional model of the Subpolar North Atlantic NNATL12.
There are known automated eddy identification methods that are widely used as a tool for studying eddy activity in statistically significant samples. The most commonly used local extrema search method depends strongly on a number of parameters chosen by an expert exploiting this approach. In order to alleviate the subjectivity issue, we first implemented the automatic identification scheme for IRs based on the local extrema search. We optimized the scheme employing Bayesian optimization framework resulting in optimal values of the hyperparameters of this eddy identification algorithm. While the optimization significantly improved the quality of the identification, we found that there is a room for further improvement of IRs detection.
As a promising alternative to the heuristic local extrema search algorithm, we propose using artificial neural networks. In this study, we employed a convolutional neural networks similar to U-Net which we trained to segmemnt the eddies. We first pretrained it on the results of heuristic IR detection algorithm. We then further trained it on the expert-labeled IRs. The resulting IR detection quality is high enough to further implement tracking algorithms.
The application of artificial neural networks, specifically convolutional neural networks akin to U-Net, has demonstrated considerable potential in enhancing the detection of Irminger Rings in high-resolution oceanic simulations. By leveraging a two-stage training process, initially utilizing heuristic algorithm results followed by expert-labeled IRs, we achieved a detection accuracy that surpasses traditional methods, thus providing a robust foundation for subsequent eddy tracking endeavors.
The integration of machine learning techniques with traditional oceanographic methodologies holds significant promise for advancing the precision and reliability of IR detection and tracking. This approach not only mitigates the subjectivity inherent in parameter selection for heuristic methods but also capitalizes on the adaptability and learning capabilities of neural networks. As such, this method presents a substantial improvement over existing techniques and contributes to a more nuanced understanding of mesoscale eddy dynamics and their influence on deep convection processes in the Labrador Sea.
Future work will focus on both IR tracking and refining the neural network model by exploring different architectures and training strategies to enhance detection accuracy further.
There are known automated eddy identification methods that are widely used as a tool for studying eddy activity in statistically significant samples. The most commonly used local extrema search method depends strongly on a number of parameters chosen by an expert exploiting this approach. In order to alleviate the subjectivity issue, we first implemented the automatic identification scheme for IRs based on the local extrema search. We optimized the scheme employing Bayesian optimization framework resulting in optimal values of the hyperparameters of this eddy identification algorithm. While the optimization significantly improved the quality of the identification, we found that there is a room for further improvement of IRs detection.
As a promising alternative to the heuristic local extrema search algorithm, we propose using artificial neural networks. In this study, we employed a convolutional neural networks similar to U-Net which we trained to segmemnt the eddies. We first pretrained it on the results of heuristic IR detection algorithm. We then further trained it on the expert-labeled IRs. The resulting IR detection quality is high enough to further implement tracking algorithms.
The application of artificial neural networks, specifically convolutional neural networks akin to U-Net, has demonstrated considerable potential in enhancing the detection of Irminger Rings in high-resolution oceanic simulations. By leveraging a two-stage training process, initially utilizing heuristic algorithm results followed by expert-labeled IRs, we achieved a detection accuracy that surpasses traditional methods, thus providing a robust foundation for subsequent eddy tracking endeavors.
The integration of machine learning techniques with traditional oceanographic methodologies holds significant promise for advancing the precision and reliability of IR detection and tracking. This approach not only mitigates the subjectivity inherent in parameter selection for heuristic methods but also capitalizes on the adaptability and learning capabilities of neural networks. As such, this method presents a substantial improvement over existing techniques and contributes to a more nuanced understanding of mesoscale eddy dynamics and their influence on deep convection processes in the Labrador Sea.
Future work will focus on both IR tracking and refining the neural network model by exploring different architectures and training strategies to enhance detection accuracy further.