Deep learning prediction of two-dimensional ocean dynamics with wavelet-compressed data.

Area of Science:

• Oceanography
• Deep Learning
• Data Compression

Background:

• Deep learning models struggle with large-scale, high-resolution ocean dynamics prediction.
• A previous divide and conquer approach improved parallel processing but introduced non-dynamical errors.

Purpose of the Study:

• To investigate wavelet compression as a method to enhance deep learning-based ocean dynamics prediction.
• To evaluate the effectiveness of wavelet compression against the divide and conquer approach.

Main Methods:

• Utilized wavelet transform to compress oceanographic datasets, reducing spatial resolution by a factor of two per compression level.
• Applied deep learning models to predict ocean dynamics using compressed data.
• Compared prediction accuracy with a previously used divide and conquer strategy.

Main Results:

• Level 3 wavelet compression, despite information loss, yielded improved predictions of 2D ocean data.
• The wavelet compression method outperformed the divide and conquer approach in predicting sea surface height.

Conclusions:

• Wavelet compression is a viable technique for improving the accuracy of deep learning models in ocean dynamics prediction.
• This approach offers a promising alternative for handling high-resolution oceanographic datasets, particularly for features like the Gulf of Mexico's Loop Current.