Earth observation is getting an agent layer that picks what to watch

Picking what to observe

Earth observation is getting an agent layer, and the first job of that layer is choosing what to observe when conditions are imperfect. In a preprint, the CRISP agentic protocol increases task success from 62.4 percent to 78.6 percent while reducing invalid tool calls from 21.7 percent to 6.1 percent. The authors also report that CRISP achieves 0.78 cloud-gap coverage compared with 0.51 for entropy-based criteria under an equal acquisition budget. Cloud cover is the practical constraint here: an agent that cannot see through cloud must choose observations that still fill the gaps in coverage. m:…4362

In the same preprint, CRISP improves EarthNet2021 surface forecasting performance by 7.7 percent on out-of-distribution splits and 8.1 percent on extreme splits. Those splits matter because they test the system where conditions differ from training, not only where they match. The reported gains therefore describe robustness under change, which is the setting where an observation agent would actually be trusted to act. m:…4362

Grounding models in evidence

A second preprint addresses a different failure: vision-language models for Earth observation that treat every token as equally informative. In a preprint, standard multimodal alignment training treats all tokens as equally informative, which is ineffective for Earth-observation imagery. The GIST framework derives token importance from evidence maps built from spectral indices and multi-scale saliency, so importance follows the geophysical evidence in the image rather than being spread evenly. m:…0562

In the same preprint, the GRACE component reduces hallucinations by replacing constant margins with evidence-consistency scores and penalizing unsupported entity mentions, while the TERRA component improves model robustness by conditioning importance heads on descriptors like biome, sensor, season, and resolution. The design is explicit about where a claim should come from: an entity mentioned without evidence in the imagery is penalized, and importance shifts with the sensor and season rather than staying fixed. For executives, the relevant point is that the model is being asked to show its grounding, not only to produce a fluent answer. m:…0562

Forecasting and scheduling by risk

A third preprint turns to groundwater, where the problem is forecasting at wells the model has never been trained on. In a preprint, foundation models adapted with low-rank adapters can accurately forecast groundwater levels at unseen wells without per-well training. GeoAI-GWFM-KR outperforms per-well LSTM networks in groundwater level forecasting at 7, 30, and 90-day horizons, and physics-informed loss functions did not improve the accuracy of per-well LSTM networks in this study. The authors report a foundation model adapted to the Korean monitoring network, so the result is tied to that network and those horizons. m:…5182

A fourth preprint schedules sensors by risk rather than by a fixed timetable. In a preprint, risk-priced scheduling reduces geohazard missed-alarm rates from 2.71 percent to 0.96 percent, and the proposed framework maintains over 89 percent performance under new climate zones, sensors, and seasons. The authors also report that the SIV-Net, PRISM, and TANDEM framework achieves 98.91 percent accuracy and 82.4 mean intersection over union on combined SMAP, MSL, Weather, EuroSAT, and FloodNet datasets. Scheduling by risk is the operational idea: spend observation effort where a missed alarm would cost the most, and keep performance when the zone, sensor, or season changes. m:…5824

Taken together, these preprints describe an agent layer that does three things: choose the next observation, ground its statements in image evidence, and allocate sensor time by risk. This suggests the value of Earth observation systems will increasingly depend on those decisions, not only on sensor resolution, since a better sensor still wastes its pass if it looks at the wrong place or reports an unsupported claim. All four results are preprints, so they should be treated as author-reported findings awaiting further validation before they guide operational deployment. m:…4362 m:…0562 m:…5824