Earth Systems Lab 2026: 3D Clouds for Climate Extremes

Project Background

Predicting the intensity and path of tropical cyclones is a major challenge in meteorology. The devastating impacts of storms such as Hurricane Dorian in 2019, Hurricane Milton in 2024 and Hurricane Melissa in 2025, which rapidly intensified from Category 1 to Category 5 superstorms, highlights the critical need for more accurate forecasting. One key reason for this difficulty is that the crucial microphysical properties of cyclones, such as ice and water content, are inadequately represented in forecasting systems. Most satellite data only provides a two-dimensional view of cloud tops, so this lack of detailed information makes it difficult to understand the complex processes that drive cyclone intensification and to provide timely warnings to affected populations. In particular, challenges remain where winds increase more than 35mph (56kph) within a 24-hour period, and as the below graphic demonstrates, this attribute is becoming an increasing trend in hurricanes as climate change advances.

Building on Earth observation systems such as CloudSat, launched in 2006, recent years have seen great advancements in the study of climate change and weather forecasting, with the launch of EarthCARE (Earth Cloud Aerosol and Radiation Explorer) in 2024. This latest satellite is equipped with active and passive sensors that provide previously unavailable data of clouds and their processes across the globe. EarthCARE provides a global operational cloud product every 16 days, derived by merging its multi-spectral imagery with observations from its cloud profiling radar.

Following the EarthCARE launch, there is increased interest in combining active sensor data with imaging to extend the value of these measurements and gain scientific insight. The team sought to build on last years ESL 2024 3D CLOUDS USING MULTI-SENSORS study to showcase how machine learning and AI can be used to derive 3D cloud products from 2D imaging observations. They would create continuous, near global 3D cloud maps as a novel dataset for research - encompassing geometric structure and microphysical properties in three-dimensions. The ultimate aim being to compare their results with EarthCAREs output to assess the projects accuracy.

Project Results

The team found that the SatMAE model without metadata performed better than the SWinSatMAE model, especially after fine tuning with 100 epochs, hypothesising that the smaller attention window of the SWin transformer was not sufficient to capture long-range correlations. The increased accuracy of the ViT SatMAE model proved worthwhile even considering the longer training times required (SWinMAE - 7 hours / SatMAE - 1 day for 50 epochs; Swin MAE 4 days / SatMAE 10 days for 100 epochs).

The below diagram shows spatial distributions of radar reflectivity (top), ice water content (middle), and droplet radius (bottom) predictions for the SWinSatMAE model (left) and the SatMAE model (middle) with the difference (right).

The best performing model, SatMAE (middle) is then shown against the baseline U-Net model (left) in the diagram below, indicating far superior performance.

Using these results the team considered the case of Hurricane Dorian. It formed in the North Atlantic in August 2019 and was forecast to remain weak as it tracked towards the Caribbean, however it underwent rapid intensification over a few days making landfall in the Bahamas as a Category 5 storm. It caused an estimated $4.3 billion in damage and resulted in 67 fatalities. There were only two CloudSat overpasses of Dorian before it intensified, leading to forecasters underestimating the development of the storm. The team's model proved that the continuous monitoring of geostationary satellite data could cover the gaps in CloudSat's coverage. The diagram below shows 2D and 3D reconstructions of Hurricane Dorian's (a) radar reflectivity, (b) ice water content, and (c) effective radius by the SWinSatMAE model. The geostationary image from GOES channel 7 is shown under each generated 3D render, with the location of the CloudSat track marked in red.

Reconstructions from this data were then used to make a 3D prediction of the hurricane at its peak intensity, as shown below as a slice though the storm centre.

All the reconstruction data was favourably comparable to the CloudSat and U-Net retrievals from the time.