Small data: 20CRv2c after compression into a 100-dimensional latent space

Small data: Near-surface temperature, wind, and MSLP, from the Twentieth Century Reanalyis, after compression into a 100-dimensional latent space

20CRv2c is on a 2 degree grid, so this set of four surface variables has a state vector of size 180*90*4=64,800: we need that many data points, every 6-hours to make the analysis video. The autoencoder compresses that 64,800-dimensional state vector into a 100-dimensional latent space, and then expands it out again. This video shows the reanalysis after this compression. So it’s the same as the original 20CRv2c video, except that it uses only 0.15% as much data to represent the weather state. As well as the weather state, the video shows the associated latent-space vector (the compressed data form) as the grid of 100 numbers at the bottom left.

• Get the data
• Train the autoencoder
• Plot a single frame
• Plot all the frames and make a video