Twenty One Percent of the Ocean Got Darker and Almost Nobody Noticed

Seventy Five Million Square Kilometres

Stand on any beach and look out. The water holds the colour it has always held. Whatever happened to it over twenty years happened at a few centimetres of lost clarity a year, spread across an area whose edge you cannot see, at a depth you cannot reach. No photograph of it exists. None could be taken.

Thomas Davies and Tim Smyth worked around that problem. Davies studies light pollution in the sea, as a marine ecologist at the University of Plymouth. Smyth, an optical oceanographer at Plymouth Marine Laboratory a short walk away, spends his working life on instruments that measure how sunlight and moonlight behave underwater. Between them they held the two halves of a question. A satellite record of ocean optics reaches back to 2002. Nobody had run a global trend through it and asked whether the sunlit layer of the sea is thinning.

They ran it. Global Change Biology published the answer in May 2025: 21 percent of the global ocean measurably darkened between 2003 and 2022 [1], an area of 75,341,181 square kilometres, quoted to the individual square kilometre because a pixel count comes out that way. Half the land surface of the planet. More than four times Russia.

Over the same period, 10 percent of the ocean got clearer. Nothing here is a uniform global dimming. The imbalance runs about two to one, and the imbalance carries the headline.

Depth carries the biological weight. Light in water stops at no wall. It fades, and somewhere in the fade sits a level below which an organism can no longer do what light lets it do: photosynthesise, hunt, see a predator coming, register that the moon has risen. Davies and Smyth name the layer above that level the photic zone. Its floor came up by more than 50 metres across 32,449,129 km², and by more than 100 metres across 9,392,219 km².

Fifty metres matches a fifteen-storey building. Across an area the size of Africa the ceiling of the dark rose by that much. No instrument on any shoreline registered it.

One Number Called Kd, and the Equation It Sits In

Everything in the paper runs through one quantity. Take a moment over it first.

Two processes take sunlight apart once it enters seawater. Water molecules and dissolved substances absorb photons outright, as heat. Particles scatter photons sideways instead, and a share of those wander back up through the surface and out of the sea altogether. Downward light at depth \(z\) therefore runs weaker than light at the surface, and the diffuse attenuation coefficient, written \(K_d\) and carried in units of inverse metres, says how much weaker.

The satellite works at one colour, 490 nanometres, a blue-green that reaches further into clear seawater than any other visible wavelength (pure water absorbs red almost at once and ultraviolet not much more slowly, so a question about how deep light gets is a question about the colour that gets deepest). The product carries the name \(K_d(490)\). MODIS Aqua, a NASA instrument in polar orbit since 2002, measures none of it directly; the instrument sees the spectrum of light leaving the top of the sea, and an algorithm turns the blue-to-green ratio in that spectrum into \(K_d(490)\) [4, 5]. A retrieval, then, rather than a reading. Every criticism of the paper starts there.

A rough feel for the numbers. The clearest subtropical gyre water on the planet sits near \(K_d = 0.02\ \mathrm{m^{-1}}\). Average open ocean runs 0.04 to 0.06. A shelf sea in summer might reach 0.12. A coastal bloom, 0.35. The mouth of a big muddy river can exceed 1.0. Fifty-fold, across water that all looks like water.

The paper's whole result: that number has crept upward across a fifth of the sea.

The equation doing all the work

Light attenuation in water belongs to a small class of environmental problems whose governing equation is both simple and correct. Irradiance falls off geometrically with depth:

Here \(I_0\) is the irradiance just below the surface. Each additional metre removes the same fraction of whatever light made it that far. The curve never reaches zero. The sea has a gloom rather than a floor.

Set the surviving fraction to one percent, the oldest convention in the field, and solve for depth:

That single line is our whole model. Plug in \(K_d = 0.02\) for 230 metres of sunlit water. Plug in 0.20 for 23 metres. Plug in 1.0, a river plume, for 4.6 metres, roughly the depth at which your own feet go out of sight.

Figure 1 draws that equation four times. Percentage of surface light runs up a logarithmic vertical axis, which turns every exponential decay into a straight line whose steepness is exactly \(K_d\), so that the four waters differ from one another by slope alone. Slope compounds.

Plain Arithmetic

No argument in this section. The arithmetic, stated.

Area where \(K_d(490)\) rose: 75,341,181 km². Area where it fell: 37,269,515 km². Ratio: 2.02 to 1. Global ocean area taken as 361,900,000 km². Risen fraction 20.82 percent. Fallen fraction 10.30 percent.

Photic depth shoaled by more than 10 m over 68,402,842 km². By more than 50 m over 32,449,129 km². By more than 100 m over 9,392,219 km². Those bands nest. Un-nested, they give three disjoint bands: 35,953,713 km² lost between 10 and 50 m, 23,056,910 km² lost between 50 and 100 m, 9,392,219 km² lost more than 100 m.

Euphotic depth at the one percent level is 4.60517 divided by \(K_d\). At \(K_d = 0.02\), 230.3 m. At 0.04, 115.1 m. At 0.06, 76.8 m. At 0.08, 57.6 m. At 0.20, 23.0 m. At 0.35, 13.2 m. At 1.00, 4.6 m.

Light remaining at 50 m: 36.788 percent at \(K_d = 0.02\), 13.534 percent at 0.04, 1.832 percent at 0.08, 0.005 percent at 0.20.

A ten percent rise in \(K_d\) removes 9.0909 percent of the euphotic depth. The result holds at \(K_d = 0.001\), at \(K_d = 10\), and everywhere between, because depth goes as \(1/K_d\) and the constant cancels.

A rise of 0.005 m^-1 removes 46.05 m at \(K_d = 0.02\), 12.79 m at 0.04, 3.39 m at 0.08, 0.56 m at 0.20, and 0.0229 m at 1.00. The first and last differ by a factor of 2,010.

Those two paragraphs are the same equation.

Why the Clear Water Is the Fragile Kind

Nothing else in the paper surprised us as much, and we spent an afternoon arguing before we believed it rather than suspecting a sign error.

Differentiate the euphotic depth with respect to the attenuation coefficient and you get

The minus sign says more murk means less depth. Obvious enough. The \(K_d^2\) in the denominator says something else: a fixed dose of murk does damage in inverse proportion to the square of the murk already present.

Put a number on it. Take the clearest water in the ocean, a subtropical gyre at \(K_d = 0.020\), and add 0.005. A small and plausible change, unremarkable on its own. The one percent depth falls from 230.3 m to 184.2 m. Forty-six metres of lit water, gone. Now take a river plume at \(K_d = 1.000\) and add the same 0.005. The one percent depth falls from 4.605 m to 4.582 m. Gone: 2.3 centimetres.

Same insult, 2,010 times the injury. The blue nowhere in the middle of the Pacific, the stretch that looks emptiest and least interesting from a boat, is optically the most delicate water on the planet.

Run the same comparison in percentage terms and the asymmetry disappears. A ten percent rise in \(K_d\) removes 9.0909 percent of the depth in the gyre, 9.0909 percent in the plume, 9.0909 percent in a puddle, the relative loss being scale free.

Which framing is right depends on what you are counting. Count metres of water column lost to an organism and the gyre holds the catastrophe; count the fraction of a light budget lost and the two waters are identical. Figure 2 carries both series together, the only honest way to plot it.

Notes From the Club Table, Three Thursdays

A White Plate on a Rope

Before satellites, before photometers, before anybody had heard of an attenuation coefficient, a priest lowered a dinner plate.

In 1865 the Papal Navy asked Angelo Secchi, astronomer to the Vatican, for a way of measuring the clarity of seawater, and he lowered a white disk on a line until it went out of sight, then wrote down the depth. The instrument has not changed in a hundred and sixty years. Build one this afternoon for the price of a tin lid and a tape measure.

Poole and Atkins, working at Plymouth in 1929 (the same town, as it happens), put a photoelectric cell in the water next to a Secchi disk and found the relationship between the two [7]:

Substitute that into the euphotic depth formula and the constants collapse into something you can carry in your head:

Whatever depth your disk vanished at, multiply by 2.7 for the level where one percent of surface light arrives. Table 1 runs that calculation across the useful range, with the metres lost to a ten percent rise in murk in the last column, and Figure 3 draws the same relation with the disagreement between published constants shaded in.

Table 1. Secchi readings converted with \(K_d = 1.7/Z_{sd}\). The last column restates the previous section in a form checkable against your own rope. Clear water loses many metres for the same relative insult. Murky water loses centimetres.

The history repays a pause. Boyce, Lewis and Worm reconstructed a century of phytoplankton change largely from archived Secchi readings, and concluded the ocean had been losing phytoplankton at around one percent of the global median per year since 1899 [8]. Critics contested that paper fiercely, on exactly the grounds that stitching old plate-on-a-rope data to modern chlorophyll measurements can manufacture a trend out of a change in method. The argument bears on this one. We come back to it.

How Much Lit Water Went Away

The paper reports areas and no volume, for a good reason. Turning an area and a depth into a habitat volume takes assumptions the satellite cannot check, and we did it anyway and labelled it accordingly, because a volume is the only form in which most people can feel the size. What follows is the club's arithmetic on Davies and Smyth's areas. Not their result. They would very likely add caveats we have not thought of.

Take the three disjoint shoaling bands. Charge each a representative shoaling depth. Multiply area by depth. Add. Table 2 holds the ledger, and Figure 4 draws the same ledger to scale, so the ink on the page runs proportional to the water.

Table 2. The volume ledger. We charge the open-ended band at its lower bound, which certainly understates it.

Three point seven five million cubic kilometres. Call it 0.28 percent of all the water in the ocean, and 5.18 percent of the epipelagic zone, the top 200 metres where nearly everything that photosynthesises lives. Spread evenly over the whole sea, a layer 10.35 metres thick. Spread over the 21 percent that darkened, 49.74 metres.

The Mediterranean Sea holds about 3.75 million cubic kilometres. On this arithmetic, one Mediterranean of lit water went in nineteen years. Per decade, 1.97 million km³. Per day, about 540 km³, or a Lake Superior every three weeks.

The only genuinely free choice in that table is where inside each band the representative depth goes, so we sampled it: 200,000 seeded draws, each band's depth drawn uniformly from its range, the open band capped arbitrarily at 150 m. Median 3,981,096 km³, with a 68 percent interval of 3,403,407 to 4,562,414 km³. Our headline figure sits at the 34th percentile of that distribution, on the low side, where we wanted it.

The Best Case Against This

We think the result is probably right. We also think it belongs to a class of results that have been wrong before, in this exact field, for reasons nobody found obvious at the time. What follows is the strongest version of the argument against it, assembled from what the authors themselves flag and from the older literature on trend detection in ocean colour.

Twenty years is not long enough, and the objection belongs to the authors before it belongs to us. The paper says so directly: twenty years is insufficient to completely rule out natural multidecadal variability [1]. No boilerplate there. Henson and colleagues worked out in 2010 how long a continuous satellite record must run before a climate-driven trend in ocean chlorophyll can be separated from the noise of interannual and decadal variability, and the answer came to about 39 years [12]. Beaulieu and colleagues reached a similar conclusion three years later, and added that the problem worsens in regions with strong decadal cycles, which covers most of them [13]. Forty years are needed. Twenty exist. The honest position is that something has been detected and cannot yet be named.

The instrument itself drifts. MODIS Aqua launched in 2002 into a sun-synchronous orbit with a nominal equator crossing time, and that crossing time has been slipping ever since, which is why NASA's Ocean Biology Processing Group flagged significant degradation in the science quality of MODIS Aqua ocean colour products from 2023 onward. The Davies and Smyth window stops at 2022, before the flagged period. Fortunate, and no proof on its own that the calibration held steady through the preceding two decades, since slow instrumental drift that mimics a slow geophysical trend is very hard to catch from inside the dataset.

Chlorophyll is not biomass, and attenuation is not chlorophyll. Behrenfeld and colleagues argued in 2016 that much of the observed relationship between warming and satellite chlorophyll is photoacclimation, cells adjusting their pigment content, rather than any change in how much phytoplankton is present [14]. \(K_d(490)\) stands a step further removed still, responding to pigment, to suspended sediment, to coloured dissolved organic matter, and to whatever else absorbs blue-green light. A real rise in \(K_d\) is a real optical fact. Biology does not follow from it automatically.

The Secchi precedent. Nature published Boyce, Lewis and Worm's century-long phytoplankton decline [8] in 2010, and it drew detailed objections, partly successful, about whether combining Secchi and chlorophyll records could generate a trend that was an artefact of the blending. Anyone reading a long optical trend in the ocean should start suspicious. That episode is why.

And a counterweight, because the case against is not the whole case. Cael and colleagues came at a related question from a different direction in 2023, taking raw remote-sensing reflectance rather than derived products, and found significant trends over a large fraction of the surface ocean, emerging faster than chlorophyll trends because reflectance is multivariate and some wavebands run quieter year to year [9]. Their own figure then came down from 56 percent to about 40 percent in a 2024 author correction [10], which cautions against first numbers and shows the field policing itself. Two independent analyses of the same satellite era, using different quantities, both find widespread change. No proof in that. More than nothing.

Our position, for what a school club's position is worth: the direction is probably real, the magnitude is uncertain at the tens-of-percent level, and the twenty-year window is the weakest joint in the whole structure. Ask again in 2035.

Who Lives in the Part That Went Away

Suppose the change is real. What lives in those 3.75 million cubic kilometres?

Start with what has to be there. Photosynthesis needs light, and the depth at which a phytoplankton cell's photosynthesis exactly pays for its own respiration is the compensation depth. The one percent convention proxies for it, roughly. Banse argued in 2004 that the field should have dropped the convention decades earlier, because the actual compensation depth depends on the species, the temperature and the day, and can sit anywhere from the 0.1 percent level to the 10 percent level [15]. Marra and colleagues revisited the question with modern data and reached a similar conclusion [16]. Every number in this article that uses the one percent level inherits that looseness.

Now go dimmer, because the paper does. Davies and Smyth set their threshold at the light level where Calanus copepods, small crustaceans that feed most of the North Atlantic, respond behaviourally. Båtnes and colleagues measured it during the Arctic polar night at around 0.027 microwatts per square metre at 490 nm [11], far below anything a human eye registers as light.

The threshold matters for what it triggers. Every evening, across the whole ocean, an enormous quantity of animal life rises from depth to feed in the surface layer under cover of darkness, and sinks again before dawn. The largest movement of biomass on Earth, twice a day. Light is the cue, by its rate of change and its absolute level rather than by its presence. Last and colleagues showed that through the Arctic winter, with no sun for months, zooplankton run this migration on moonlight, to a lunar rather than a daily rhythm [17].

Raise the depth of that faint cue by 50 metres and the animals following it rise too, into a shallower, warmer, more crowded layer, nearer to visual predators, competing for the same food. Nobody has measured the consequences, because the shoaling was only described last year.

A precedent exists at smaller scale. Aksnes and colleagues documented darkening in Norwegian fjords and argued it had already driven regime shifts in the mesopelagic communities living there [18], and Opdal and colleagues found a centennial decline in North Sea water clarity large enough to delay the spring phytoplankton bloom [19]. Local systems, local causes, and the open ocean is a different animal. They do establish that darkening water changes who lives in it, the step the global paper cannot yet take.

What Would Settle It

Three things, and one of them is already happening.

The first is time. The Henson threshold of roughly forty years estimates how long the noise takes to average down rather than stating a rule of nature, and the order of magnitude is right. MODIS Aqua's record began in 2002. Add SeaWiFS before it and Sentinel-3 and PACE after it, cross-calibrated carefully, and a forty-year series comes within the working lifetime of anybody currently in a school science club. Cross-calibration is the hard part, and somebody's full-time job.

The second is in-water validation at depth. Every number in the field comes from an algorithm turning the colour of the sea surface into a property of the water column, and Argo floats now carry optical sensors, so a float profiling irradiance directly can check \(K_d\) retrieved from above against \(K_d\) measured from inside. Enough of those, over enough years, and the calibration-drift objection loses most of its force.

The third is attribution, the genuinely difficult one. Knowing that a fifth of the ocean darkened says nothing about why. The paper's candidates are sediment and nutrient loading near coasts, and changes in circulation and stratification offshore, and those imply completely different futures. Decisions made on land can in principle reduce coastal loading. A circulation change answers to nothing.

Meanwhile the result stands as the first global measurement of something everybody had assumed static, the open ocean having served until now as an optical constant, a background against which other changes were measured. The background moves. It has a trend, and the trend points down.

What We Would Tell Somebody Who Read Only the Headline

Four things.

The ocean did not get 21 percent darker. Twenty-one percent of the ocean got measurably darker, by an amount that varies from place to place and stays small in most of them. Ten percent got clearer over the same period. The story is the imbalance, roughly two to one, spread across an area the size of half the world's land.

"Darker" describes a coefficient, not the view from a boat. Swim in the middle of the darkened region in 2003 and again in 2022 and you would notice nothing. The change lives at depth, in the last few percent of the light, in a layer you would need a rope and a photometer to find.

The volume figure in this article, 3.75 million cubic kilometres, is ours and not the paper's, and it comes from multiplying their published areas by representative depths we chose, which is where the uncertainty sits. We picked conservative values, published the code, and put a 68 percent interval on it, so treat the result as a way of feeling the size of the reported areas rather than a measurement of anything.

And the last one, the reason this became a field note. Two people found a planetary-scale change by running a trend through a public dataset that anyone could download. The satellite had been recording it since 2002. The data were free. Nobody had asked. The lesson about where the remaining discoveries sit is not a comfortable one.