Neural Compression for Cloud Analytics

Clouds visually indicate the atmospheric state of our Planet that is driven by complex dynamics. Today’s global weather models generate km-scale forecasts at sub-minute resolution such as in the nextGEMS. Comprehending the vast volume of information requires intelligent compression methods.

We share progress on how E2S methods enable the coupling of atmospheric models and Earth observation data through the utility of embeddings generated by deep neural networks. We demonstrate how atmospheric embeddings help predict land cover, how they guide the generation of meso-scale cloud patterns, and how Earth observation embeddings help to characterize the statistics of clouds over the past 25 years.

Read More