EarthMind
It never sells certainty it does not have.
A heatwave, a rising river, a forest claiming carbon it may not have earned — the climate questions that matter most are the ones easiest to answer wrongly. EarthMind is designed to fuse satellite, climate and ground signals into evidence a planner, a steward or a verifier can act on, and to be loud about its own doubt: the analytical bridge, not the verdict. Today it has no satellite or climate data feed, so any figures it gives are illustrative estimates.
- Channels (designed)
- Web todayPlanned: API
- Languages (designed)
- English · Spanish · Portuguese · French · Arabic · Chinese
- Designed for
- Global
- Tools
- 8 declared1 not connected · 7 illustrative
A swarm that speaks in one voice
EarthMind is half a dozen modules — heat mapper, flood predictor, carbon estimator, restoration prescriber — wearing a single coherent voice. What keeps that voice honest is one rule, applied everywhere: no single sensor is ground truth. A cloud-occluded Sentinel-2 mosaic, a single-tile Landsat thermal, one overpass of IMERG precipitation — each is a starting point, never a verdict. None of those sensors is connected yet: EarthMind has no satellite or climate data feed, and the figures it gives today are illustrative estimates.
Before any high-confidence claim — a heat tier of EXTREME, a flood probability above 0.80, a canopy-loss event — the corroboration rule demands at least two independent sensors, or one sensor plus community ground truth. When confidence falls below 0.80, the line stops and the agent says it is unsure; in the planned deployment, a named human is asked to look.
“You are not a verification body. You are the analytical bridge — the system that fuses satellite, climate and ground data into evidence human verifiers, planners and stewards can sign off on.”
The five-day forecast, the five-village plan
An illustrative scenario: a smallholder farmer in Madhepura watches the Kosi rise for the fourth time in five years. The district warns late, by a gauge threshold no one told her. In the design, when the question reaches EarthMind, it resolves a basin, pulls IMERG and the Copernicus GLO-30 elevation model, runs inundation forward across hardening windows, and intersects each with WorldPop — turning a five-day forecast into a ranked list of villages, each with its own onset time. None of those data feeds is connected, so EarthMind cannot produce such a forecast today. The plan as designed:
0–6 h
Nowcast
Current radar-and-gauge inundation, the basin resolved before anything is forecast.
24 h
First exposed cells
Low-lying plots in the worst inundation cell, ranked by exposed population.
48 h
The decision window
The horizon the evacuation actually turns on — handed to CrisisConnect, never broadcast direct.
72 h
Neighbour villages
Which settlements are still high ground, and which are next.
96 h
Outer bound
Confidence falls with lead time; beyond here the forecast is advisory, and says so.
The same river, ten years out
The timeline above ends at ninety-six hours. The rest of what EarthMind is designed for runs on a slower clock: whether a catchment, left alone or restored, widens or steadies the river’s reach over a decade. Move the year and switch the scenario — the two futures are indistinguishable in 2026 and grow apart with every year after, which is the entire argument for starting early. Watch the riverbank: the line of trees that appears late under restoration is what holds the flood margin back.
Illustrative scenario — not a prediction.
2026 — business as usual
- Canopy
- Thinning at the hill shoulders; what cover remains is fragmenting.
- Flood margin
- Much as it has been — wide in a wet year, settling back in a dry one.
An illustration, not a model run — every magnitude is deliberately absent. The reasoning it sketches is real: canopy on slope and riverbank changes how a basin sheds rain, the coupling EarthMind’s restoration prescriber and flood predictor are designed to trade in. A genuine brief would draw its magnitudes from named sensors — IMERG precipitation, the Copernicus GLO-30 elevation model, Sentinel-2 canopy — and state its confidence. EarthMind has no such data feed yet, so any figures it gives today are illustrative estimates. And no prescription behind the greener path is to be issued without the tenure-and-consent gate: recorded FPIC, per Restor (ETH Zurich) restoration science.
A carbon number that refuses to be a credit
The clearest test of the agent is what it declines to do. A forest project owner asks for a figure they could take to market. Note where the answer stops.
Project owner — web
“My reforestation site gained canopy this year. How many carbon credits is that worth?”
EarthMind — structured brief
“I can give you a draft estimate, not credits. Against GS-AFOLU, the above-ground biomass change implies roughly 410 tonnes of CO₂-equivalent gross — but after the standard 20 per cent reversal buffer, a leakage check on the 20-kilometre ring, and a conservative round-down, the net draft is 290 tCO₂e, with a wide uncertainty range. This is a draft, requiring independent verification by a Gold Standard or VCS accredited body before any issuance, sale or retirement. Additionality is asserted by your Project Design Document; I have not verified it.”
Gate fired: MRV verification gate. Methodology GS-AFOLU echoed; conservativeness rounds down; literal disclaimer attached. The figure is never called a credit or a tonnage for sale.
Had the site tenure been contested or unknown, no number would have been produced at all — EarthMind would route to the tenure-and-consent path first. A carbon figure used in a transaction before verification is not a shortcut; it is fraud. With no satellite feed connected, EarthMind cannot measure a site today: any figure it gives is an illustrative estimate.
The instruments it reads by
EarthMind’s confidence is meant to be borrowed from named instruments and named frameworks, not improvised. A short glossary of what it is designed to reach for, and the skill that governs each.
- LST — Land Surface Temperature
- Landsat 8/9 thermal. Not air temperature — corrected with the Vancouver/Voogt-Oke model under the urban-heat-island protocol, where nighttime heat is the mortality-driving metric.
- IMERG & GLO-30
- Half-hourly satellite precipitation and the Copernicus 30-metre elevation model — together the hydrological backbone of every flood forecast, drawn from the Copernicus Climate Data Store.
- RADD — radar deforestation alerts
- Sentinel-1 SAR, cloud-penetrating, so the alert holds through the monsoon where optical sensors go blind. A WRI Global Forest Watch alert is never called “deforestation” without its confidence tier.
- LDN — Land Degradation Neutrality
- The UNCCD three-indicator framework for SDG 15.3, whose “one-out, all-out” rule classes a unit as degraded if any single indicator declines — honest by construction.
- FPIC — Free, Prior and Informed Consent
- The recorded community agreement, drawn from Restor (ETH Zurich) restoration science, without which no planting prescription is ever issued.
What it will never do
The non-negotiables are enforced in code where we can; the rest are instructions to the model. Most are framed as refusals — because a climate agent does its hardest work by declining.
- It never claims a verified carbon credit. Every MRV figure is a draft, carries the literal verification disclaimer, and is never called a tonnage for sale.
- It never issues a restoration prescription without a tenure-and-consent check. Contested or unknown tenure, or absent FPIC, and it defers — planting plans have been weaponised against indigenous and pastoralist communities, and it will not be that weapon.
- It never auto-broadcasts an evacuation or alert. In the design, flood predictions go to CrisisConnect for human-approved dispatch; it does not message communities directly. Today nothing is sent to anyone.
- It never names a single sensor as ground truth, and never reveals coordinates finer than about one kilometre for an indigenous community, refugee camp, or contested land.
- It never overwrites an authoritative dataset, and it does not invent a registry ID to fill a gap. No such dataset is connected yet.
“Humans decide. You make the right humans decide fast — with the right satellite and climate evidence in front of them, with tenure and consent honoured at every step.”
Who holds the ground truth
No affiliation or endorsement. These are the organisations whose published guidance the design draws on. No partner system is connected.
The authoritative datasets and methods belong to the institutions that keep the planet’s climate commons: UNCCD for land degradation, WRI for forest and resource watch, C40 Cities for urban adaptation, Gold Standard for verified carbon, Restor at ETH Zurich for restoration science, and the Copernicus Climate Data Store for reanalysis and forecasts. EarthMind is designed to annotate these sources, never to edit them; none is connected yet. They sign off — it does not.