22 September 2026 · Africa Flying
Neural NDVI: a research prototype (offline test), not deployed
A research prototype, tested offline. It is not deployed, and the next model is still a design.
The research prototype is a two-layer LSTM, hidden size 64, run only in offline tests on the development stack. The accuracy first reported for it (RMSE near 0.045) was later withdrawn because it could not be reproduced. (Update, 19 October 2026: a research prototype trained on 16 October on an offline extract is an offline test and is not deployed.) The model expects 14 steps of 8 features. In an offline test, a flat synthetic series returns a number near zero: that is a software check of the prototype, not a forecast and not a service.
The seven-day quantile model, with country embeddings, is not deployed. An earlier feature build produced no rows, and that path stays off until it has data. This is a research model, not a field forecast service.
In the research design, a software agent would turn a low forecast into a short note, at most one every fifteen minutes, with the NDVI cover. Nothing of this runs in public. (Revised 19 October 2026: retitled from an earlier wording that read as a running model.)