An agent that fine-tunes a geospatial foundation model for site-to-grid evapotranspiration, GPP and transpiration whenever FLUXNET, HLS and the climate covariates have all advanced, then checks whether the new model is actually better before publishing anything.
Now:
The design, animated with example data. It plays on its own; pause, or click a step to hold it.
An evapotranspiration model built by fine-tuning a geospatial foundation model. The pretrained encoder supplies the representation of surface state from HLS reflectance; a multi-output head maps that representation, together with climate covariates, onto three tower-derived targets: evapotranspiration, gross primary productivity and transpiration. The agent wrapped around it watches the input archives and fine-tunes when new data arrives, so the published product tracks the observational record instead of freezing at whatever was available on the day it was built.
Towers measure total evapotranspiration, not transpiration. The agent derives tower-level transpiration two ways — the Transpiration Estimation Algorithm (TEA; Nelson et al., 2018) and the underlying water use efficiency method (uWUE; Zhou et al., 2016) — and keeps both. Where they agree, the label is used as is; where they diverge, the spread is carried into training as label uncertainty rather than averaged away.
Both methods draw on tower GPP, so the transpiration and GPP labels are not independent of each other. The comparison step scores the three outputs separately for that reason, and a gain in transpiration is not counted as confirmation of a gain in GPP.
FLUXNET gives on the order of a few hundred CONUS site-years, which is small for the dimensionality of the problem. A pretrained encoder carries structure learned from far more imagery than the flux record contains, so each cycle adapts a head rather than relearning what a vegetated surface looks like. It also makes continuous retraining affordable: a fine-tune is hours, a full pretrain is not.
Any one stream alone adds version churn without adding information, so the gate requires all three.
If any check fails, the incumbent stays in production and the challenger is kept as a candidate with its diagnostics.
The model estimates land surface fluxes. Its transpiration output is only as good as the partitioning behind its labels, and no tower measures transpiration directly. It also does not see water leaving through a cooling tower, a pipe, or any other non-land-surface pathway, and no amount of resolution changes that. Anything built on top of this for water accounting needs an independent demand estimate.
This page is a demo so collaborators can see the workflow. Production runs on an HPC cluster: watchers on a scheduled job, fine-tuning and upscaling submitted through the batch scheduler, and this interface served as a static page from the run directory.
| Name | What it does | Size |
|---|
Listing is illustrative. Watchers and the gate run as a scheduled job; fine-tuning and upscaling are submitted to the batch scheduler and write into a run directory that this page reads.