Who gets looked at

In 2010 a magnitude 7.0 earthquake struck New Zealand. Nobody died. In 2010 a magnitude 7.0 earthquake struck Haiti. The death toll ran to six figures.

Same hazard. Same size. Two orders of magnitude between the outcomes, and none of that difference is geology.

That comparison is Roger Bilham and Nicholas Ambraseys's, and the paper it opens is called Corruption kills. Their finding, published in Nature in 2011, is that 83 per cent of all deaths from building collapse in earthquakes over the preceding thirty years happened in countries that are anomalously corrupt for their income level. Substandard materials, codes not enforced, buildings put where buildings should not go.

The same shape holds across every hazard type, not just the ones that shake. UNDRR's mortality analysis for 1996 to 2015 found that people living in disaster-affected areas of low-income countries died at about 130 per million, against 18 per million in high-income countries. Seven times. Higher-income countries went through 56 per cent of the disasters and took 32 per cent of the deaths; lower-income countries went through 44 per cent and took 68 per cent.

And it holds within countries, not only between them. A team led by Sara Lindersson at Uppsala took geocoded mortality from 573 major floods across 67 countries between 1990 and 2018 and found that the wider the income gap inside a country, the higher its flood mortality, after controlling for GDP per capita and for how many people the water actually reached. Inequality was not standing in for poverty. The shape of the distribution was doing work on its own.

Why that happens is not mysterious

Think about what an unequal place looks like physically. Housing gets sorted. The land nobody wants, the floodplain, the unstable slope, the fault scarp, the strip beside the drainage channel, is where the cheapest housing goes, because it is the cheapest land. Then the infrastructure follows the money rather than the risk, so the drainage, the seismic retrofit, the early warning, the road you evacuate on, all cluster where they are least needed.

None of that is a disaster-response failure. It is a set of decisions made over decades, and the earthquake or the cyclone or the flood is just the moment the ledger gets read out loud.

Now the part that made me want to write this down

If you wanted to fix any of that, the first thing you would need is to know where those households are. Not roughly. Specifically.

That is exactly what satellite-derived poverty maps are supposed to do, and they are used for it: allocating humanitarian aid, targeting cash transfers, deciding which districts get a programme. The method is reasonable. Take a modest amount of ground-truth survey data, train a model on the imagery over those surveyed places, then predict wealth everywhere else, because the imagery is everywhere and the surveys are not.

Emily Aiken, Esther Rolf and Joshua Blumenstock audited that pipeline across ten countries, and the finding is quietly devastating. These maps are good at telling urban from rural. They are considerably worse at distinguishing wealth within urban areas or within rural ones.

The model has learned the difference between a city and a village. It has not really learned the difference between two villages.

And the difference between two villages is the entire question. Nobody needs an algorithm to tell them that the countryside is poorer than the capital. What a targeting system is for is picking, among places that look broadly alike from orbit, which one has the unreinforced housing, which one is downstream of the moraine, which one is about to lose its harvest. That is the resolution where the errors live.

Where the survey data is, and is not

The thing that makes this self-reinforcing is where ground truth comes from. Survey coverage is not sprinkled evenly over a country. It follows roads, follows cities, follows wherever the last funded round of fieldwork happened to reach, and skips the places that are expensive or dangerous or politically inconvenient to get to.

Which means the model is trained on a sample that already under-represents the hardest-to-reach places, then asked to make its most consequential predictions about exactly those places. The gap in the training data becomes a gap in the map, and the gap in the map becomes a gap in the aid.

Nobody chose that. There is no meeting where somebody decided that remote settlements should be scored badly. It is an artefact of where it was convenient to collect data, and it is now sitting inside systems that move real money.

A place with no record is not a place with no problem

This is the sentence I keep coming back to, and it is the reason Caelus exists at all.

Colonias along the Rio Grande are unincorporated. That is an administrative fact, and it has a physical consequence: being outside every service boundary also means being outside the boundary of whoever would install the gauge. So the water comes, and the satellite records it, and the ground record stays empty, and the next time anyone asks whether that place floods, the honest answer from the files is that there is no evidence it does.

The same structure shows up everywhere once you start looking for it. A valley in Rasuwa where the flood that arrived in August 2026 registered on seismometers on other continents before anyone downstream knew. Flares over the Niger Delta counted nightly from orbit since 2012, with nobody joining that count to who lives within a kilometre of one. Groundwater under the Thar measured from space for two decades while the land above it was recorded as greening. None of these events were unmeasured. They were unaddressed, which is a different failure with a different fix.

Absence of data reads as absence of need, and that is the specific lie this whole field is built on top of.

So what do you actually do about it

Three things, none of which are glamorous.

• Work at the resolution where the error is. A national estimate is not wrong, it is just useless for deciding which of six colonias to reach first, or which block of a damaged district to send an assessor to. Structure-by-structure, parcel-by-parcel analysis is more boring and more expensive and it is the only thing that answers the question that was actually asked.
• Treat model output as a hypothesis, not a finding. Where somebody who lives in a place says our inference is wrong, they are right and we are wrong. Ground truth from a resident outranks a prediction every single time, and building the correction path first is what separates a useful map from a confident one.
• Deliver to the people already there. A finding that reaches a dashboard nobody in the district can open has not been delivered. It has been published, which is a different verb and a much easier one.

None of this requires better satellites. The satellites are fine. Sentinel-1 goes over all of these places on a fixed schedule and sends back radar that works through cloud and darkness, free, forever, whether or not anyone is reading it.

The bottleneck was never the instrument. It is that reading the data carefully for a place with no budget, no gauge, and no constituency is nobody's job. That is a choice, and it is one small enough to change.