Neural Nets Shrink the Planet: Introducing LGSMs

Neural nets are unparalleled at compressing very very large datasets. It occurs to me that one of the fairly untapped application is storing entire Earth in 3D in a model (Large Geo-Spatial Model or LGSMs?).

Some estimates done in ~15 mins:

• Using current compression ratio, whole Earth model would probably need just ~1T param model (1px/m^2).
• About 100 small airplanes for a year to scan whole Earth (using basic trig).
• It would roughly cost $200M to acquire this data and likely same amount to train a model.

Applications of this model could be beyond VR and gaming. You can query this model with all kind of obscure questions such tree cover, receding glaciers, unexplored climbing routes, scenic driving roads, potential mines, panoramic views. Temporal/seasonal data even more magical.

Model can then also be used to answer what-if questions, make predictions and generative purposes. It could be interesting to see what kind of emergence happens in geo-spatial domain.

Discussion