Global atmospheric data assimilation with multi-modal masked autoencoders

TL;DR

It is shown that EarthNet performs a form of data assimilation producing a global 0.16 degree reanalysis dataset of 3D atmospheric temperature and humidity at a fraction of the time compared to operational systems and that the resulting reanalysis dataset reproduces climatology by evaluating a 1 hour forecast background state against observations.

Abstract

Global data assimilation enables weather forecasting at all scales and provides valuable data for studying the Earth system. However, the computational demands of physics-based algorithms used in operational systems limits the volume and diversity of observations that are assimilated. Here, we present "EarthNet", a multi-modal foundation model for data assimilation that learns to predict a global gap-filled atmospheric state solely from satellite observations. EarthNet is trained as a masked autoencoder that ingests a 12 hour sequence of observations and learns to fill missing data from other sensors. We show that EarthNet performs a form of data assimilation producing a global 0.16 degree reanalysis dataset of 3D atmospheric temperature and humidity at a fraction of the time compared to operational systems. It is shown that the resulting reanalysis dataset reproduces climatology by evaluating a 1 hour forecast background state against observations. We also show that our 3D humidity predictions outperform MERRA-2 and ERA5 reanalyses by 10% to 60% between the middle troposphere and lower stratosphere (5 to 20 km altitude) and our 3D temperature and humidity are statistically equivalent to the Microwave integrated Retrieval System (MiRS) observations at nearly every level of the atmosphere. Our results indicate significant promise in using EarthNet for high-frequency data assimilation and global weather forecasting.

Figures (9)

• Figure 1: EarthNet ingests multi-dimensional Earth observations from varying orbits and spectra. Sensor modalities in the first four rows have 12 hours of input sequence with a number of channels. The last row is a static elevation variable defining the topography. Sub-images are extracted spatially of size $(144, 144)$ with a token size of $(16, 16)$. Tokens are encoded with a vision transformer and embedded per sensor modality. After tokens pass through the backbone transformer, each decoder sees all context tokens. This process is applied as a moving window across the image and reassembled using Hann windows.
• Figure 2: Background departures of EarthNet's temperature (a-f) and specific humidity (g-l) against observations temporally averaged across the spatial and vertical dimensions. The top row shows MiRS average temperature and humidity values across February and March 2024. EarthNet's 1 hour background state average temperature and humidity are shown in the second row. The third row shows error departures as computed from 1 hour background state predictions minus the MiRS observation.
• Figure 3: Sensitivity analysis and sensor importance measured by relative mean absolute errors over the land and ocean. (Top row) Each sensor is dropped individually while reconstructing all modalities and comparing errors to the baseline of all modalities included. (Bottom row) One sensor is taken as input to reconstruct all with errors computed as above. The analysis is split into the land (left) and ocean (right) regions to delineate surface types.
• Figure 4: Comparison with radiosonde soundings shows that EarthNet's performance is similar to MiRS observations in matching radiosonde observations of temperature (a) and humidity (b). EarthNet outperforms MERRA-2 and ERA5 reanalyses for humidity predictions between 50 and 500 hPa, without the benefit of having assimilated radiosonde data. Scatter plots show a strong linear relationship between ERA5/MERRA-2 reanalysis and radiosonde temperature observations (c). Relative humidity scatter plots show more normally distributed errors for EarthNet and MiRS at high humidity values (d).
• Figure 5: Analysis of EarthNet's hourly performance shows minimized error during the middle of the non-overlapping 12-hour periods used for inference.