RoadSense Nigeria · measured

A satellite can show you a bad road. It could not show you a road getting worse.

A cracked, waterlogged street in Lagos: a bus and several vans nose through standing water while people walk along the edge of the flood.
A flooded, rutted road in Lagos, 2020: buses and vans picking their way through, pedestrians on foot.Photo: Tunde Olowookere, CC BY-SA 4.0
An empty multi-lane dual carriageway with fresh surfacing, painted markings, street lights and a grass verge, under a wide sky.
The same country: a dual carriageway outside Abuja, empty, lit and drained.Photo: Taini Ozeto, CC BY-SA 4.0

Only 30 per cent of Nigeria's roads are paved, according to the World Bank. The rest are earth, gravel, or something a lorry leaves behind in the rain. The two photographs above are the ends of that range: a Lagos street where buses and vans nose through standing water, and a dual carriageway outside Abuja, empty, lit and drained.

A photograph tells you which of those two a road is on the day somebody pressed the shutter. It cannot tell you which way the road is going. The flooded street may flood every rainy season and be no worse than it was ten years ago. The clean carriageway may be quietly losing its verge, year by year, in a way no single picture would ever show.

In 2017 the World Bank put 40 per cent of federal roads in poor condition. That is one year's reading, not a trend, and it says nothing about any particular road. What a maintenance budget needs to know is different: of the 850 stretches of federal highway measured here — 18,419 km of road, each stretch long enough to have a name and a budget line — which ones are getting worse?

What the record says

Workers in high-visibility vests laying a new concrete slab by hand beside a narrow residential street, with buildings and a parked car behind.
Concrete being laid in Bogije, Lagos: what the money buys when it arrives.Photo: Ayodele Adeniji, CC BY-SA 4.0

Nigeria's procurement record is the only national paper trail of road spending. It names a road, a state, a kind of work, an amount and a date: money allocated.

It does not say whether the road got better. A contract records a promise, and a commissioning ceremony is a photograph of a promise, nothing more. The money may have reached the road, part of it, or none of it.

The record shows where money moved; satellite imagery shows what the ground did. The test below asks whether the second predicts the first.

The result

Can a free satellite photograph tell a road agency which of its roads are falling apart? This page reports the attempt: 850 stretches of Nigerian federal highway, 22 months, and one test that was written down before it was run. The test failed.

0.3901the score. Half, 0.50, is a coin toss

Plausible range
0.3196 to 0.4671
Bar to clear
0.60 point estimate, with the lower end above 0.50
Direction
below a coin toss (0.50): roads with recorded spending score as less degraded than roads without
Sample
644 roads compared, 66 of them with a payment on record, 578 without

The number is a ranking score, and it is worth saying plainly what it ranks. Take two roads, one with a payment on record and one without. A score of 0.50 means the measure picks between them no better than a coin. Above that, it usually picks the funded road as the one in worse shape. Below it, it picks the funded road as the one in better shape — which is what happened here. So the failure is not "the effect is small". It is "the measure points the wrong way", and a measure that points the wrong way cannot be used to send crews anywhere.

The five figures

The map: movement, not condition

Across 850 corridors over 10 months, 69 fell, 347 rose and 434 stayed within 0.01 of zero either way. The outer band starts at 0.02, the ninetieth percentile of absolute movement.

Where the gap between the road and its verge moved, corridor by corridorMap of 850 federal road corridors across Nigeria, coloured by mean year-over-year change in the gap between the road surface and the vegetation beside it, over 10 months. 27 corridors fell by 0.02 or more, 42 fell by 0.01 to 0.02, 434 stayed within 0.01 of zero, 202 rose by 0.01 to 0.02 and 145 rose by 0.02 or more. 0 corridors carry no usable change value and are drawn in neutral grey. 48 corridors have a dotted outline because their value rests on fewer than 10 monthly values. This is movement, not condition: the colour says how much a corridor moved against its own history and against the national month, not how good or bad the road is now, and it cannot rank roads against each other nationally.

This shows movement, not condition. A corridor that has always been poor and one that has always been good look the same here. It cannot rank roads against each other nationally.

Shows: the mean year-over-year change in the gap between the road and its verge for each corridor, over the 10 months where a year-earlier pair exists (2025-09 to 2026-06). The comparison is each corridor against itself, the same calendar month one year earlier, so a road's own level is differenced out and the season is held fixed: a corridor that did not move sits at zero by construction, and the grey centre band reads “no change against that reference”, never “a good road”. 0 corridors have no usable change value and are drawn in neutral grey rather than given a colour; 48 rest on fewer than 10 monthly values and carry a dotted outline. 63 corridor-months inside the window carry no value and are left out, not interpolated (183 across the whole panel). Does not show: current condition, or that a dark violet corridor is in worse shape today than a dark blue one — only that its contrast moved further in one direction over the period.
The same numbers as a table
Corridors by change in the gap between the road and its verge, over the change window
BandCorridors
contrast falling by 0.02 or more27
contrast falling by 0.01 to 0.0242
steady, within 0.01 either way434
contrast rising by 0.01 to 0.02202
contrast rising by 0.02 or more145
no usable change value0

The map is one SVG drawn from 127,091 source vertices, simplified to 110,003 for this page at a tolerance of about half a display pixel, and clipped to the corridor bounding box (11.8 degrees of longitude, 9.1 of latitude).

Seasonal fall in the gap between the road and its vergeLine chart of the national monthly mean contrast over twenty-two months. It rises to a peak of 0.1330 and falls to 0.0647, dropping 38.3 per cent between 2024-09 and 2024-12.Mean road-minus-verge NDVI contrast, national, by month0.060.090.140.1160.071falls 38.3%2024-092025-012025-052025-092026-012026-05
The seasonal trap. Shows: the raw gap falls 38.3 per cent on corridor means over the full 22 months, from 2024-09 to 2024-12, with no road changing (F22 measured 42 per cent on tile-row means over the first four months). The two measures answer the same question on different rows and are not interchangeable. Does not show: that any corridor deteriorated, or that the road itself changed.
The same numbers as a table
National mean of road-minus-verge NDVI contrast, by month
MonthMean contrast
2024-090.1156
2024-100.0951
2024-110.0814
2024-120.0713
2025-010.0647
2025-020.0708
2025-030.0863
2025-040.0965
2025-050.1023
2025-060.1121
2025-070.1243
2025-080.1330
2025-090.1235
2025-100.1075
2025-110.0914
2025-120.0814
2026-010.0751
2026-020.0785
2026-030.0902
2026-040.1017
2026-050.1101
2026-060.1233
The first test score collapses onto a single valueHistogram of the score over 850 corridors. 726 corridors share one value, 0.000241, the spread inside that group being floating-point noise at 7.9e-18; every other bar is a sliver.Corridors by value of the first test's score0850726 (85.4)-0.00710.0015Score: minus the mean two-way residual, per corridor
The broken instrument. Shows: 726 of 850 corridors (85.4) share one and the same value, 0.000241, and the spread inside that group is floating-point noise (7.9e-18). This is the score the first test used; it could not have detected anything, because a score that takes one value on most of the sample has no ranking to give. A gate now refuses to run an area under the curve before checking that the score can vary at all. Does not show: that the three instruments in the next chart are broken; they are continuous, and that gate passed for all three.
The same numbers as a table
Corridors by value of the first test's score, the most common values first
Score valueCorridors
0.000241726
-0.00105460
-0.00100817
-0.00291212
-0.0024317
every other value28

Distinct values at full precision: 168 across 850 scored corridors; histogram drawn with 20 equal bins between -0.0071 and 0.0015.

Three instruments against the chance line and the barForest plot with a point and a ninety-five per cent interval for each instrument. All three points and all three intervals lie to the left of both the chance line and the bar.clears the bar0.200.300.400.500.600.70chance 0.50bar 0.60Primary / year-over-yearSecondary / 15-month trendReference / raw levelAUC, higher is better; the violet line is the bar
The two tests, and the reference. Shows: every interval lies entirely below the chance line and below the bar of 0.60, and the second test's interval reaches above chance only by a fraction of a point. Does not show: an effect size of zero, or that the imagery carries no signal — with 66 positive corridors the intervals are wide and observational.
The same numbers as a table
Area under the curve, matched sample, with intervals
InstrumentAUCIntervalAnalysedLabelledUnlabelledPermutation p
Primary / year-over-year0.39010.3196 to 0.4671644665780.00
Secondary / 15-month trend0.43180.3544 to 0.5137644665780.08
Reference / raw level0.31330.2461 to 0.3881644665780.00
Where the variance sitsSingle stacked bar: corridor 81.2 per cent, month 5.9 per cent, residual 12.8 per cent.Share of the raw variance, additive corridor plus month81.2%5.9%12.8%Corridor 81.2% | Month 5.9% | Residual 12.8%Month entered first, then corridor; 183 corridor-months missing.
Where the variation sits. Shows: with month entered first and corridor second, 5.9 per cent of the raw variance is between months, 81.2 per cent between corridors, leaving 12.8 per cent as interaction and noise — the only part a change measure could use. The split is order-dependent: entering corridor first gives 82.1 per cent corridor and 5.1 per cent month, which is why the order is stated rather than implied. Does not show: that the residual is condition; land cover, weather and registration differences all sit in it.
The same numbers as a table
Where the variation sits in the raw gap, month counted before road
ComponentShare of total
Month, entered first5.9%
Corridor, entered second81.2%
Residual12.8%
Corridor first instead82.1%
Month second instead5.1%
Corridor-months used18,517

How the measurement was made

For each corridor and each month, the mean vegetation index of a band along the carriageway is subtracted from the mean of the verge beside it, at 10 metre resolution, from Sentinel-2 via the DE Africa platform. That gives one value per corridor per month: 18,700 corridor-months over 22 months, with 183 left blank rather than filled.

What was checked before the test ran
CheckResult
Imagery months present22 of 22, 0 missing
Imagery rows, corridors, duplicate keys69,212, 850, 0
Weather rows, missing, nulls18,700, 0, 0
Score varies, all three instruments850 distinct values each, 0.1% on the modal value
Score is not a proxy for coveragecorrelation -0.0123, AUC 0.5089
Decision rule, written before the test0.60 point estimate, lower bound above 0.50

The test asks whether the contrast falls faster on corridors with a procurement record than on comparable corridors without one. Corridors are matched exactly on road class and on length band within class; the statistic is the probability that a labelled corridor ranks as more degraded than an unlabelled one, with a bootstrap interval and a permutation null. The bar was written down before the test ran, and was not lowered after it failed.

What this does not claim

Four things this does not claim.

  1. Movement, not condition. The score has no level: a corridor that was always bad and one that never changed can produce the same value.
  2. Federal corridors only. The network is the named federal highways; feeder roads, which carry much of the traffic, are not in it.
  3. Reanalysis, not gauges. The rainfall covariate is a gridded reanalysis product, not a rain gauge, and it is context rather than a measured cause.
  4. Observational, not causal. 66 labelled corridors is a small, non-random, spending-driven sample. A procurement record is evidence that money was allocated, never that a road failed.
An unpaved earth road running through a busy market street: shop fronts and umbrellas on both sides, a lorry ahead, people walking across the surface.
An earth road through a market in Akara: no surface, no drainage, and a town's traffic on it.Photo: Okechukwu Iroegbu, CC BY-SA 4.0

Where the numbers come from

Every figure on this page is emitted by the script that built it and re-checked on each build against the file it came from; a number typed by hand would fail the build. The code, data manifests and findings register are on GitHub.

Credits