Augmenting the EnKF with a Shallow Convolutional Neural Network
Poster, International Symposium on Data Assimilation, Fort Collins, CO
High-resolution datasets are often too computationally expensive to assimilate into operational forecasts using standard methods. We demonstrated in a proof-of-concept experiment that machine learning methods may be able to augment traditional approaches to improve predictive performance in such a situation. Using Lorenz-96 as a test system, we used the EnKF to assimilate synthetic observations and trained a shallow CNN to reproduce the EnKF results. Then observations alternated between all variables observed (representing high resolution data) and half of the variables observed (representing low resolution data). Compared with ignoring the high-resolution data and using the EnKF on the low-resolution data only, using the trained CNN on the high-resolution data improved the accuracy significantly.