Aaron Zuspan @aazuspan.dev · Dec 8 1 / 2
For comparison, here's the actual and predicted modern imagery from MODIS. A climate-only model definitely misses some topographic and soil detail, but seems to capture land cover surprisingly well.
3 likes 2 replies
?
Replies
Now let’s see the hyperreal Earth, with 2×predicted - ground truth. (I’m kidding. This is great.)
The model is a stupidly simple pixel-based random forest trained with a grid of sampled points (below) using downscaled bioclimatic predictors derived from temp and precip (via WorldClim) and stretched MODIS RGB. I also tried predicting in HSV and LAB space, but RGB looked the best 🤷♂️