Limitations¶
What this tool cannot tell you, and where it is likely to be wrong.
There are no accuracy numbers¶
We have no labelled ground truth, so we do not publish an accuracy, and neither should you infer one. Every figure in these docs is an engineering measurement — tile counts, timings, file sizes, coverage extents — all real and re-runnable, none of them accuracy.
If you need real numbers for your area, validate against an independent dated source. For forestry in British Columbia the obvious candidate is the province's published harvest polygons.
For this reason, we provide Sentinel-2 imagery basemaps so you can independently validate the quality of your outputs. Have fun and share any interesting findings!
A high change score is not proof the land changed¶
The embedding summarises a whole year, climate and phenology included. A drought, a late spring or an unusual snow year moves the embedding without anything on the ground being cut, built or burned. As Burns (2026) puts it, large interannual differences can arise even when the land surface itself has not changed (Google Earth blog).
The score cannot separate those causes. A cutoff finds deviation, and deciding which deviations are land-surface change is what the labelling step is for.
Annual only — and late-year change is muted¶
An embedding is a summary over a fixed period, and the published product's period is the calendar year (AlphaEarth Foundations). Two consequences:
- Nothing dates a change more precisely than the year it happened.
- A change late in year B reads weaker than it should. Clear a block in October and the year-B embedding still summarises nine months of standing forest, so the score is diluted and the object may fall below your cutoff. If you suspect a late-year change, compare against the following year instead.
The same averaging absorbs short-lived events. Floods, and fire scars that green up within the year, are poorly served; permanent change is what this suits.
The first class you label accepts the opposite change¶
With one class there is nothing to discriminate against, so objects are ranked by similarity to the average of your examples — and that average is dominated by what the place was. Label three cleared blocks and regrowth from the same forest baseline scores just as well.
On a synthetic fixture — planted clusters, not real ground — 100% of the reverse change was accepted, identically in all three feature modes. It is the first state every session is in. Label a second, contrasting class as early as you can — that is what switches the head to logistic regression.
Classes do not travel well¶
The model receives no coordinate information, which its authors speculate is why it needs far more examples to learn climate gradients (AlphaEarth Foundations). Expect a classifier trained in one climatic zone to (possibly) degrade in another. Sometimes it works pretty well though!
Coverage is not truly global¶
Nothing exists beyond ±83.36°, nothing over open ocean, and coverage grew between 2017 and 2025 — so an early year paired with a recent one covers less ground than either does alone. See Data coverage for the map and the numbers.
The dimensions mean nothing on their own¶
The 64 numbers are not bands and have no individual interpretation. You cannot inspect why two years differ, only that they do, and by how much. Interpretability and independent validation of these embeddings are both open problems.
The reference photos are composites¶
The before/after imagery is cloud-free annual compositing, not a dated acquisition. It answers "what was here that year" and never "what date did this change". It is also non-commercial — see Reference.