Assessing soil health where no soil maps exist — using satellite embeddings as a stopgap. Sound approach or fooling ourselves?
A lot of agricultural land — especially newly assessed, leased, or smallholder fields — has no usable soil survey data at all. No lab history, no maps, nothing to base management decisions on.
We've been exploring whether satellite-derived embeddings can give a first, continuous estimate of soil properties in exactly these data-poor situations. The idea isn't to replace lab sampling, but to provide a baseline where the alternative is literally no data.
The method: thousands of georeferenced lab samples from across Europe, paired with Satellite embeddings (64-dimensional annual "fingerprints" per pixel), trained with Random Forest to estimate pH, organic carbon, CaCO₃ and macronutrients continuously across a field.
Where I'd love this community's input:
- For assessing basic soil health (OC, pH) on an unmapped field, is a remote-sensing estimate genuinely better than nothing — or does a wrong-but-confident number do more harm than an honest "unknown"?
- How much of an annual embedding's OC signal is soil vs. vegetation/management confounding?
- Which properties would you flat-out refuse to estimate this way?
Not claiming this solves anything — more interested in where practitioners think the honest limits are for field scouting (agri-chat.de)
