# U-Net Kalman Filter (UNetKF): An Example of Machine Learning-assisted Ensemble Data Assimilation

arXiv: https://arxiv.org/abs/2403.12366 · HTML: https://mlsanity.com/paper/2403.12366/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 7 of its 40 references that are in the map.
- First posted on arXiv 19 Mar 2024.

## Builds on (most cited first, 7 of 7)

- [U-Net: Convolutional Networks for Biomedical Image Segmentation](https://mlsanity.com/paper/1505.04597/) (2015)
- [MetNet: A Neural Weather Model for Precipitation Forecasting](https://mlsanity.com/paper/2003.12140/) (2020)
- [Improving data-driven global weather prediction using deep convolutional neural networks on a cubed sphere](https://mlsanity.com/paper/2003.11927/) (2020)
- [Using machine learning to correct model error in data assimilation and forecast applications](https://mlsanity.com/paper/2010.12605/) (2020)
- [Learning Variational Data Assimilation Models and Solvers](https://mlsanity.com/paper/2007.12941/) (2020)
- [Integrating Recurrent Neural Networks with Data Assimilation for Scalable Data-Driven State Estimation](https://mlsanity.com/paper/2109.12269/) (2021)
- [Analog ensemble data assimilation and a method for constructing analogs with variational autoencoders](https://mlsanity.com/paper/2006.01101/) (2020)

## Abstract (arXiv)

Machine learning techniques have seen a tremendous rise in popularity in weather and climate sciences. Data assimilation (DA), which combines observations and numerical models, has great potential to incorporate machine learning and artificial intelligence (ML/AI) techniques. In this paper, we use U-Net, a type of convolutional neutral network (CNN), to predict the localized ensemble covariances for the Ensemble Kalman Filter (EnKF) algorithm. Using a 2-layer quasi-geostrophic model, U-Nets are trained using data from EnKF DA experiments. The trained U-Nets are then used to predict the flow-dependent localized error covariance matrices in U-Net Kalman Filter (UNetKF) experiments, which are compared to traditional 3-dimensional variational (3DVar), ensemble 3DVar (En3DVar) and EnKF methods. The performance of UNetKF can match or exceed that of 3DVar, En3DVar or EnKF. We also demonstrate that trained U-Nets can be transferred to a higher-resolution model for UNetKF implementation, which again performs competitively to 3DVar and EnKF, particularly for small ensemble sizes.