The Extent of Doubt

WATCHING GETS CHEAPER

More eyes

Sputnik 1 did not observe Earth. On October 4, 1957, it crossed the sky as a polished metal sphere with a radio beacon inside it. People heard the beep before they understood the new architecture around them. The first artificial satellite was less an eye than a sign: the planet had acquired an outside.

Landsat 1 arrived in 1972 with multispectral instruments, and a different kind of watching began. A field, a forest, a coastline, a city edge could be revisited from orbit. Later came radar that could see through cloud and darkness, hyperspectral sensors that could separate narrow bands of reflected light, CubeSats small enough to ride into orbit in groups, and commercial fleets that made repeat imaging feel ordinary.

The scale still catches me. Thousands of machines now move above us, and a growing share of them are pointed back at the ground. Planet has built fleets that can image Earth's land surface at daily cadence. Radar operators can revisit a floodplain through monsoon cloud. Hyperspectral instruments can ask whether a pixel carries a chemical signature rather than only a color.

It is tempting to call this progress and stop there. But more eyes do not automatically produce a better mind. They produce more chances to notice, and also more chances to confuse noticing with knowing. A daily image of a field tells me that the field changed. It does not, by itself, tell me what the field is becoming.

THE HARDER HALF

Seeing is not knowing

Judea Pearl's ladder of causation stays in my head because it gives language to a discomfort I keep meeting in Earth observation. Association asks what tends to appear with what. Intervention asks what would happen if something changed. Counterfactual reasoning asks what would have happened if the world had turned slightly differently.

Most satellite systems are very good at the first kind of question. A thermal anomaly appears; an active fire is likely. Water spreads across pixels that were dry yesterday; flooding has probably begun. Vegetation reflectance drops; stress may be present. These are not trivial achievements. They save time, money, and sometimes lives.

But the question that keeps bothering me usually arrives before the visible evidence. Which dry forest is about to burn? Which basin is becoming dangerous before the water crosses the road? Which crop is entering failure while it still looks green from a distance?

A three-hour warning and a three-day warning can look like versions of the same product, but they are not the same thought. One begins after the world has started confessing. The other asks the model to infer a pressure building inside the system. That is where the facts stop feeling like facts and start feeling like bets.

A sensor can tell me something changed. The harder question is whether it can tell me what kind of future that change belongs to.

A satellite can see a large wildfire. That is impressive, but it is also late. By the time heat becomes bright enough to stand out from orbit, something is already burning.

The useful question is harder. Can we look at a landscape on Monday and say that one small patch is unusually likely to burn on Thursday? No flame. No smoke. Just fuel, heat, wind, moisture, terrain and human activity waiting in an uncertain combination.

This is the difference between seeing an event and understanding the conditions that make an event possible.

Fire detection has improved because the whole chain has improved. Sensors revisit more often. Ground stations receive data faster. Processing runs automatically. Alerts can reach emergency teams without waiting for a human analyst to inspect every image.

In August 2026, ESA and Eumetsat said they had shortened delivery of some Sentinel-3 fire detection data over Europe from the normal three-hour requirement to roughly 60 to 90 minutes. Nothing about the satellite sensor itself had to change. The gain came from changing how quickly data was sent to the ground.

That is a useful reminder. Planetary intelligence is not only about better models. Sometimes the breakthrough is removing an hour from a pipeline.

Before a fire, the landscape is full of weak signals. Vegetation may be drying. Soil moisture may be low. Hot, dry air may be pulling water from leaves. Wind may be forecast to rise. Dead fuel may have accumulated after months without enough rain.

Satellites can measure parts of this story. They can estimate vegetation condition, land surface temperature, moisture and recent change. Weather models add wind, humidity and temperature. Terrain models add slope and aspect.

But a dangerous landscape does not always burn. Most dry places on most dry days do not become disasters. Prediction has to identify the rare combination that crosses the line.

A model can know almost everything about fuel and weather and still miss the event because it does not know where the spark will come from. Lightning is one source. So are power lines, machinery, campfires, vehicles, farm burning and deliberate ignition.

This makes wildfire prediction a strange problem. Part of it is physics. Part of it is ecology. Part of it is weather. Part of it is human behaviour.

A landscape can be ready to burn for days. The event begins when possibility meets a spark.

People often imagine prediction as a clean statement: a fire will start here at 3:20 PM. That is rarely the right target. A useful system may instead say that the chance of ignition and rapid spread is much higher than normal across a small area for the next twelve hours.

That sounds less dramatic, but it is more actionable. Crews can pre-position. Utilities can inspect vulnerable lines. Authorities can restrict risky work. Cameras and satellites can increase attention over a smaller zone.

Good prediction is not about sounding certain. It is about changing the next decision.

A system that warns everywhere is technically safe and practically useless. Fire agencies cannot deploy crews to every dry hillside. Utilities cannot shut power across every windy region. People eventually stop listening to alerts that rarely matter.

The real measure of a prediction system is therefore not only how many fires it catches. It is how much attention it wastes. A warning has a cost even when nothing burns.

This is where Earth observation meets economics. The best model may be the one that helps a human decide where to spend the next hour, not the one that produces the most colourful risk map.

We are getting better at spotting fire quickly. That will save lives and land. But the deeper scientific milestone is still ahead.

It arrives when a system can watch millions of quiet hectares, ignore almost all of them, and notice the few places where fuel, weather, terrain and likely ignition are becoming dangerous together.

The satellite will not be looking for fire anymore. It will be looking for the conditions from which fire emerges.

GROUND TRUTHS

The farmer's shadow

Agriculture is where this becomes personal for me. A farmer does not need a satellite to know the sky has been strange. The body learns seasons. Soil has a smell before rain. Leaves carry small warnings. People who work with land often trust signals that would look unscientific until a model rediscovers them under another name.

NDVI made one of those signals computable. By comparing near-infrared and red reflectance, it gave chlorophyll a rough orbital shadow. EVI, red-edge indices, radar soil moisture, and missions like SMAP made the shadow more layered. A crop could be read through greenness, canopy structure, surface moisture, and stress.

NASA Harvest pushes those layered signals a step further, into a forecast. Its GEOCIF model pairs historical yield records with satellite vegetation signals and weather data, learning which combinations of temperature, rainfall, and stress tend to precede a good or bad harvest in a given region. A companion model, ARYA, follows the shape of the growing season itself, tracking how a crop's development tracks against temperature and linking that curve to the yield it eventually produces. Both convert a pixel time series into a number a food security analyst can act on months before harvest.

That sounds like understanding until the decision becomes specific. Which week should planting begin? How much water should be applied? Is this field stressed in a way that will reduce yield, or only in a way that looks alarming from space? A vegetation index can be precise and still not answer the question the grower is actually holding.

There is an old farming proverb: the best fertilizer is the farmer's shadow. I used to read that as nostalgia. Now I read it as a warning about proximity. Maybe the satellite does not replace the shadow. Maybe it gives the shadow a second angle, and we still have not learned how to combine the two without pretending one has defeated the other.

A field can be machine-readable and still not be understood.

When a leaf absorbs sunlight, not all of that energy becomes growth. A tiny fraction is released again as a faint glow called fluorescence. Human eyes cannot see it. A carefully built spectrometer can.

This matters because the glow is linked to photosynthesis, the process by which plants turn light, water and carbon dioxide into stored chemical energy. For decades, satellites have been very good at showing where vegetation is green. A new generation of measurements wants to ask a different question: how actively is that vegetation working?

A crop can remain green while already under stress. A forest canopy can look healthy in a conventional image while photosynthesis has slowed. Colour tells us something important, but it does not tell us everything happening inside the leaf.

Vegetation indices such as NDVI became powerful because they made plant cover measurable at enormous scale. They helped turn satellite images into agricultural and ecological tools. But they are mostly indirect measures.

Solar-induced fluorescence offers another window. It is faint, difficult to isolate and much closer to the machinery of photosynthesis itself.

ESA's FLEX mission is scheduled to launch with Copernicus Sentinel-3C on 15 September 2026. FLEX carries a Fluorescence Imaging Spectrometer designed to measure the weak fluorescence emitted by plants as they absorb sunlight.

The mission will work with Sentinel-3 observations of land and atmosphere. One satellite helps describe the scene. The other looks for the subtle signal coming from plant function.

The pairing is interesting because it moves Earth observation one small step from appearance toward process.

The obvious hope is earlier stress detection. If photosynthetic activity changes before leaves visibly yellow or wilt, fluorescence could reveal trouble sooner than ordinary colour imagery.

But turning a global satellite measurement into a field decision will not be automatic. Clouds, canopy structure, crop type, viewing geometry and spatial scale all matter. FLEX is expected to produce global fluorescence maps at about 300 metre resolution, useful for science but much coarser than a single small farm.

The mission should therefore be judged carefully. Not by whether the images look new, but by whether the measurement changes what people can know and do.

Photosynthesis is one of the largest flows of carbon on Earth. Plants pull carbon dioxide from the atmosphere and move carbon into leaves, wood, roots and soils. Small changes across enormous areas matter to the global carbon cycle.

That makes fluorescence interesting far beyond agriculture. It may help scientists test how vegetation responds to drought, heat and changing climate conditions. It can also help improve models of how much carbon ecosystems are taking up.

For the first time, we may begin to watch a global biological process through a signal generated by the process itself.

There will be a temptation to treat fluorescence as a direct meter of plant health. It is not that simple. Photosynthesis changes with light, water, temperature, species and the way plants regulate energy. The same fluorescence value can mean different things under different conditions.

The hard work will be calibration and interpretation. Scientists will compare the satellite signal with towers, field instruments, crop measurements and ecosystem models.

This is how new planetary measurements usually become useful. First we learn to see a signal. Then we spend years learning what the signal really means.

Much of Earth observation has been about shape, colour, temperature and motion. Those are powerful properties, but life is also process.

A leaf opens pores. Water moves. Carbon enters. Sugars form. Energy is stored. Stress changes the whole sequence. If satellites can begin observing those processes rather than only their visible consequences, the idea of planetary intelligence changes.

We stop asking only what Earth looks like from space. We begin asking what Earth is doing.

BUILT SYSTEMS

Proxies everywhere

The same pattern appears in infrastructure. A solar developer can look at years of satellite-derived irradiance before anyone has placed an instrument on the site. That is extraordinary. It turns an unknown field into a candidate before a team ever travels there.

Then the ground station arrives anyway. The modeled sunlight has to be adapted to the site before money, insurance, and engineering decisions become comfortable. I find that sequence revealing. The satellite did not remove uncertainty. It made uncertainty cheaper to approach.

Construction monitoring works in a similar way. Repeat imagery can show whether earthworks advanced, whether a roadbed moved, whether a transmission corridor cleared, whether a dam project changed shape between visits. A lender can watch many projects without sending people to all of them. The visible world becomes an audit trail.

But a bridge can look complete before it is trustworthy. A grid can look mapped before it is resilient. A methane plume can be detected by an instrument like TROPOMI, and still the larger question remains: what is the system becoming under stress? Counting what is visible is useful. It is also a very particular kind of ignorance.

Climate TRACE pushes that single detection toward a global ledger. The coalition fuses data from more than 300 satellites and over 11,000 ground, air, and sea sensors through machine learning trained to attribute emissions to individual sources, tracking upward of 70,000 power plants, steel mills, and ships by name rather than by country-level estimate. A company buying steel can now ask which specific mill produced it with the lowest emissions, a question no self-reported national inventory could ever answer at that resolution.

COUPLED SYSTEMS

Systems with memory

Climate makes the problem harder because the planet remembers. A warm year does not vanish when the calendar resets. Heat enters oceans. Ice retreats. Forests dry. Permafrost thaws. The next state depends on the path into it.

Sea ice is a simple example until it stops being simple. Bright ice reflects much of the sunlight that reaches it. Dark ocean absorbs far more. Lose ice, expose water, absorb heat, lose more ice. The albedo numbers can be written cleanly on a page, but the lived system is not clean. It has thresholds, delays, feedbacks, and local exceptions.

IceNet pushes that same feedback loop into a forecast. Built by the British Antarctic Survey and the Alan Turing Institute, it is trained on thousands of years of climate simulations alongside decades of real observations, then predicts monthly sea ice concentration up to six months ahead at 25-kilometer resolution. In seasonal forecasts of summer sea ice, especially the extreme low years that matter most for shipping routes and wildlife, it has outperformed ECMWF's SEAS5, the physics-based model built to do exactly this.

GRACE and GRACE-FO made mass loss visible in a way ordinary images could not. They sensed tiny changes in gravity between paired spacecraft and turned ice sheets into measurements of weight. That is one of the most beautiful ideas in modern Earth science: weighing ice from orbit by feeling gravity shift.

And still I do not know what it means to understand the thing being weighed. If Greenland loses mass this year, what does that mean for a fishing season, a monsoon, an insurance model, a coastal city, a farmer deciding whether the old calendar can still be trusted? The measurement is planetary. The consequences are local. Something difficult lives in the conversion.

All models are wrong, but some are useful.George Box, statistician, 1978

FROM SIGNAL TO INFERENCE

Models that guess

AI makes the question sharper because it is very good at producing the feeling of understanding. A model can fuse multispectral imagery, radar, elevation, weather, labels, and historical events into an answer that arrives faster than any human analyst could produce one. Sometimes the answer is useful. Sometimes it is only fluent.

Foundation models trained on large Earth-observation archives have already shown real promise. They can transfer to flood mapping, land-cover classification, crop analysis, and change detection with less labeled data than older workflows needed. That matters. Sample efficiency is not decoration; it changes what small teams and under-mapped regions can attempt.

But I keep returning to the same discomfort. A model that generalizes across a benchmark may still be interpolating inside the world it has already seen. Fire risk in a new climate regime, crop stress under a strange monsoon, flood behavior after a city changes its drainage, these are not just new images. They are altered mechanisms.

This is where causal structure, physics-informed models, and invariant learning become interesting to me. Not because they sound more serious, but because they admit the weakness in pattern recognition. They ask whether the relationship still holds when the surface changes. I do not know how far that gets us. I only know that accuracy on yesterday's distribution is a fragile comfort.

We can only see a short distance ahead, but we can see plenty there that needs to be done.Alan Turing, 1950

A motion-triggered camera in a forest does not know what it is photographing. It only knows something moved, so it fires, whether that something is a jaguar or a branch swaying in the wind.

Multiply that camera by the thousands now deployed across protected areas, running for months between visits, and the bottleneck stops being detection. It becomes review. A single research network can accumulate millions of images in one season, the large majority of them blank triggers: wind, shifting light, an animal that had already left the frame before the shutter caught it.

Ecologists have always accepted that trade. A camera never tires and never scares an animal off by being there, but every frame it produces still has to pass in front of a trained eye before it becomes usable data. For decades, that eye, not the camera, was the actual limit on how much of a landscape could be watched at all. The same limit shows up in a different form with audio recorders left running in the field, and the shift worth paying attention to is not that a machine can suddenly notice something a person could not. It is that the machine can get through all of it, which nobody could previously afford to do.

Google's SpeciesNet is an image classifier, a model trained to sort a photograph into one of a fixed set of categories, built on an architecture called EfficientNet V2 M and trained on more than 65 million camera-trap images pooled from the Wildlife Insights research community and public archives. It sorts into more than 2,000 labels: a species where the image supports one, a broader group such as felidae or mammalia when it does not, and non-animal categories like blank or vehicle for the frames that were never wildlife to begin with.

The honest numbers matter more than the headline. Google reports that the model detects an animal's presence in 99.4% of the photos that actually contain one, reaches an identification at the species level 83% of the time, and gets 94.5% of those species-level calls right. Multiply it through and roughly one photo in five still needs a person to finish the identification. That is not a small gap, but it is a very different problem from the one it replaced: a researcher at Wake Forest University used the model to work through an 11 million photo backlog in a matter of days, and a camera network in Ecuador scaled to 446 cameras and more than 100,000 images in a single year after adopting it, coverage neither project could have reviewed by hand at that pace.

What changed is not that the model replaced the ecologist. It re-routes their attention. The four confident calls out of five get sorted out of the queue, which leaves the finite number of trained reviewers free to spend their time on the contested fifth, and on the rarer species the training data barely covers in the first place. A backlog that used to mean data nobody had gotten to yet now means data a person only has to check, not first discover.

A passive acoustic recorder left running in a forest or reef for weeks produces audio no team could review at listening speed and keep the rest of their job. Google DeepMind's Perch was trained to classify nearly 15,000 species from sound, mostly birds, along with frogs, crickets, grasshoppers, and some mammals, and to generate embeddings: a compressed numerical fingerprint of a sound clip that other tools can compare against each other, useful for tasks well beyond the species label alone.

Hawaii's honeycreepers, native forest birds under severe pressure from avian malaria and habitat loss, several species already extinct and others down to a few hundred individuals, are exactly the kind of case where review speed decides how much ground a small team can cover. The University of Hawaii's LOHE Lab used Perch to identify honeycreeper calls in field recordings nearly 50 times faster than the conventional method of scanning spectrograms by eye, letting the same handful of researchers monitor more sites than they could ever have covered by ear alone.

The clearest case for using a model at all shows up underwater. Perch's updated version was trained on essentially no underwater audio, yet the same embeddings transfer well enough to flag meaningful events in coral reef soundscapes, because a healthy reef and a degraded one sound different, in the density of snapping shrimp and fish calls, before either condition is visible to a diver. No dive team or hydrophone crew was ever going to log that continuously across a reef system. The recorder already could. Until a model could sort what it captured, that continuous record had no way to become usable data at all.

Population science needs more than a count of animals seen. It needs to know whether the whale shark in this month's photo is the same one photographed off the same coast last year, which is the difference between a real population estimate and a number that just counts sightings twice. The nonprofit Wild Me built Wildbook, now branded Sharkbook for sharks, around exactly that problem: the pattern of spots behind a whale shark's gills is as individual as a fingerprint, and the platform matches a new photo against a global library using an approach built the way face-matching software is, only trained on skin patterns instead of faces, returning researchers a ranked list of likely matches rather than a single automatic verdict.

The system extends past the photos researchers submit on purpose. It checks public video uploaded and tagged as whale shark footage, reads the description, runs the same matching model to check for a known individual, and uses location data to filter out anything clearly filmed in an aquarium. A tourist's holiday clip becomes an unplanned entry in a global mark-recapture study, the same kind of repeat-sighting data biologists once relied on physical tags for, without the tourist ever finding out.

None of this replaces the biologist's judgment. A ranked list of candidate matches still needs a person to confirm the top one, and the whole system depends on people continuing to photograph and upload the animals in the first place. What it changes is scale: years of manual photo-matching become weeks, across a population no research team could ever physically travel to in full.

A species label, a matched call, a confirmed individual: all three systems are doing the same job at a scale no person could match. They turn an unmanageable pile of raw evidence, photographs, hours of audio, uploaded video, into a much smaller pile a trained ecologist can actually look at.

None of them explain why a population is falling or a habitat is failing. A species identification is still only a better observation, not an account of the mechanism behind a decline. That harder question, the one this volume keeps returning to, still belongs to the person deciding what the cleared backlog actually means.

A model that finally has time to look at everything still has to hand the interesting cases back to someone who knows what they are looking at.

A WORKING DEFINITION

What I mean by planetary intelligence

I do not want planetary intelligence to become another inflated phrase. If it means only more sensors, we already know how that story goes. If it means only larger models, we may only be scaling the same confusion. The phrase is useful to me only if it names a harder standard.

For now, I mean something narrow: a system that can sense change, carry some model of mechanism, update across domains, and support decisions before the event has fully announced itself. Not just flood mapping after water arrives. Not just crop stress after yield is already lost. Not just methane detection after the plume is visible. Something closer to a shared estimate of what the planet is in the middle of becoming.

The pieces exist in fragments: orbital sensors, gravimetric mass tracking, weather assimilation, causal fire models, hydrology, crop models, foundation models, human local knowledge, and the stubborn ground truth that keeps correcting everything from below. What I cannot yet see is the coupling. I cannot see where these fragments become trustworthy together.

That may be the first honest place to leave this note. Earth has become measurable in astonishing detail. I do not want to mistake that for comprehension. I want to stay with the harder question: what would it take for a system to understand a planet well enough that people would trust it before the proof arrives?

That's here. That's home. That's us.Carl Sagan, Pale Blue Dot, 1994