Steven Ramage

Google DeepMind now compresses a year of satellite imagery for every 10m of Earth into 64 numbers. A team just used them to find every tomato field in California. A task like detecting tomato fields and distinguishing them from wheat fields can actually be pretty annoying to do. Raw satellite imagery is a mess of clouds, shadows and shifting seasons, so getting a crop map out of it normally means hand-picking which colour bands and seasonal patterns distinguish one crop from another, then redoing much of that work when you move to a new year or a new region. A model tuned to spot tomatoes in one season loses its edge in the next. For something as ordinary as knowing which field grows what, that is a lot of custom engineering to maintain, and it is exactly the fragile step everyone wants to delete. A team at UC Davis tested a way to delete it, using AlphaEarth Foundations. Instead of handing you raw pixels, AlphaEarth pre-digests a full year of satellite observations for every 10-metre patch of ground on Earth into a compact string of 64 numbers, a kind of fingerprint of everything that happened on that patch across the year. It ships one of these fingerprints per location per year from 2017 onward. The bet the Davis team wanted to test was simple. If the hard summarising work is already done for you, a fairly ordinary model should be able to read those fingerprints and map a crop with none of the usual hand-tuning. To find out they leaned on California's 2018 statewide crop map, which told them the ground truth for thousands of fields. They took 4,742 known tomato fields and 4,742 fields growing everything else, a deliberate mix of alfalfa, wheat, corn, beans and orchards, fed the fingerprints into a standard image-segmentation network, and split the data so the fields it was tested on sat in genuinely different places from the ones it learned on. That last part matters, because testing on land next door to the training fields mostly measures memory, not skill. The scores came back almost suspiciously high. Around 99% of pixels classified correctly, an F1 of 99.0% balancing false alarms against misses, and a predicted tomato area that overlapped the true area about 98%. More telling than the headline number is where the model hesitated. They ran it a hundred times over with random pieces switched off each time, watching where its answers wobbled, and the uncertainty piled up almost entirely along field boundaries while the interiors stayed rock solid. That's the pattern you'd want to see: confident about whole fields, unsure only at the fuzzy edges where one crop meets another. As you may know, I'm a bit up in the air about these embeddings datasets. They make inference cheaper and quicker, but sometimes they're outperformed by more vanilla approaches. Anyway, as with LLMs, we're just at the beginning and they'll only improve over time. So watch this space... link to full article: arxiv.org/abs/2605.21804

Crop maps drawn by the most hyped geospatial foundation models fall apart the moment you move them to a farm they have never seen. Here is the pitch you have been sold. Take a giant model that has stared at mountains of satellite imagery, give it a little training on your task, and it will map almost anything anywhere on Earth. For farming that promise is gold. You could label which field grows which crop, watch those fields change across a season, and steer food supply with a map that updates itself. The whole appeal rests on one quiet assumption nobody had properly stress-tested. That a model taught over one stretch of land still works when you drop it somewhere with different soil, weather and farming habits. So a team at the University of California, Riverside built the fair test. They took three of the flagship geospatial foundation models, named Prithvi, SpectralGPT and SatMAE, and ran them across four wildly different states. Iowa, North Carolina, California and Minnesota. It comes down to how they split the ground. They trained on one region, tuned on a second, then tested on a third, so the final exam always landed on farmland the model had genuinely never seen. That last move is everything. Test a model on the same region it learned, and you are mostly measuring memory. Test it on new ground, and you measure whether it actually understands crops. All three buckled. Moved to unfamiliar land, they retreated to guessing only the crops that blanket the Midwest, corn and soybeans, and went blind to everything rarer. Picture a crop map of California that simply cannot see the almonds, the grapes, the lettuce. That is the failure mode here, and it is precisely the blind spot that makes a map worthless for the specialty and minor crops that carry diversity and feed people. The thing sold as general-purpose turned out to be a local expert wearing a global costume. There was a second sting. To compare the three on equal footing, the researchers had to squeeze them all into one shared input format, and that squeeze hurt each of them differently. So even the head-to-head ranking gets muddy. You cannot cleanly say one design beats another when the test rig itself bends each one out of shape, which is its own cautionary tale about how these leaderboards get built. Step back and the pattern is familiar across machine learning. A model looks brilliant on the data near where it grew up and gets dramatically worse the further you carry it from home. Farming just makes the failure vivid, because soil and season and crop mix change every few hundred miles, and a tool that only works in Iowa is not the tool anyone was promised. The fix the authors push is unglamorous and overdue. Stop grading these models on home turf. Make the standard test a deliberate move to unseen geography, so the score reflects the one thing real deployment always demands, which is travel. Grade on the same region and you flatter the model and fool the field. link to full article: arxiv.org/abs/2606.29664