Snowfall as an Analog for Noise

[Bill Ferris]

We're getting a late winter storm in Flagstaff and, while enjoying the view from my back deck, it occurred to me that snowfall is a good analog for noise in photography.

In the above photo, the picnic table on my deck represents a sensor (or film negative, or glass plate) in a camera. The falling snow represents photons falling upon the sensor.

If you averaged the depth of the snow per square foot of the table surface, that would represent the exposure (light from the scene per unit area of the sensor) used to make the photo. It's the total accumulation of all snow (all photons; all light) that's fallen upon the table that determines how much noise there is and the visibility of that noise in a photo.

As is true of photons emitted or reflected by something, there's randomness in snowfall. While snow might be falling at a different rate of 1-inch per hour in another part of Flagstaff, we've been getting about 2-inches per hour in my neighborhood. That's the rate of snowfall (scene brightness) at my back deck.

In the photo, you'll see the randomness of snowfall exists even at the picnic table level. The surface of the 6 inches of snow atop the table is uneven. That randomness in the depth of the snow, is noise.

You can see the texture in the snow because of variations in its brightness. Those subtle variations add contrast to the scene and make the pebbled texture of the snow visible. That's what noise in a photo looks like. Its randomness in the brightness of a scene that is otherwise evenly lit.

The above photo was made after it had been snowing continuously for about 3 hours. There's approximately 6 inches of snow on the picnic table. The variation in the depth of the snow - the noise - is minor in comparison with the total snowfall - the total signal.

That's how noise works. The greater the total signal (the more snow there is on the table) the smaller the noise portion of that totality will be. Just as randomness in snowfall isn't as obvious when a lot of snow has fallen, noise in a photo isn't as prominent when a lot of light is used to make that image.

If I'd made this photo after it had been snowing for an hour, the depth of the snow per square foot on the table (the exposure) would have been much less. The totality of snow on the table (the total light; the signal) would have been much less. The randomness in snow depth across the table (noise) would also have been less but would have been a greater portion of the total snowfall (signal) at that time. The randomness (noise) would have been more prominent, more obvious to the eye.

This is how noise works in photography. The more light we put on the sensor (or negative, or glass plate) the less prominent noise will be in the photo. This can be achieved by maximizing exposure, which by extension maximizes the total light used to make a photo.

--
Bill Ferris Photography
Flagstaff, AZ

 

[robgendreau]

We're getting a late winter storm in Flagstaff and, while enjoying the view from my back deck, it occurred to me that snowfall is a good analog for noise in photography.

In the above photo, the picnic table on my deck represents a sensor (or film negative, or glass plate) in a camera. The falling snow represents photons falling upon the sensor.

If you averaged the depth of the snow per square foot of the table surface, that would represent the exposure (light from the scene per unit area of the sensor) used to make the photo. It's the total accumulation of all snow (all photons; all light) that's fallen upon the table that determines how much noise there is and the visibility of that noise in a photo.

As is true of photons emitted or reflected by something, there's randomness in snowfall. While snow might be falling at a different rate of 1-inch per hour in another part of Flagstaff, we've been getting about 2-inches per hour in my neighborhood. That's the rate of snowfall (scene brightness) at my back deck.

In the photo, you'll see the randomness of snowfall exists even at the picnic table level. The surface of the 6 inches of snow atop the table is uneven. That randomness in the depth of the snow, is noise.

You can see the texture in the snow because of variations in its brightness. Those subtle variations add contrast to the scene and make the pebbled texture of the snow visible. That's what noise in a photo looks like. Its randomness in the brightness of a scene that is otherwise evenly lit.

The above photo was made after it had been snowing continuously for about 3 hours. There's approximately 6 inches of snow on the picnic table. The variation in the depth of the snow - the noise - is minor in comparison with the total snowfall - the total signal.

That's how noise works. The greater the total signal (the more snow there is on the table) the smaller the noise portion of that totality will be. Just as randomness in snowfall isn't as obvious when a lot of snow has fallen, noise in a photo isn't as prominent when a lot of light is used to make that image.

If I'd made this photo after it had been snowing for an hour, the depth of the snow per square foot on the table (the exposure) would have been much less. The totality of snow on the table (the total light; the signal) would have been much less. The randomness in snow depth across the table (noise) would also have been less but would have been a greater portion of the total snowfall (signal) at that time. The randomness (noise) would have been more prominent, more obvious to the eye.

This is how noise works in photography. The more light we put on the sensor (or negative, or glass plate) the less prominent noise will be in the photo. This can be achieved by maximizing exposure, which by extension maximizes the total light used to make a photo.

Good explanation; I like it better than the buckets in a field in a rainstorm idea.

And is the bench the smaller sensor?

 

[Bill Ferris]

We're getting a late winter storm in Flagstaff and, while enjoying the view from my back deck, it occurred to me that snowfall is a good analog for noise in photography.

In the above photo, the picnic table on my deck represents a sensor (or film negative, or glass plate) in a camera. The falling snow represents photons falling upon the sensor.

If you averaged the depth of the snow per square foot of the table surface, that would represent the exposure (light from the scene per unit area of the sensor) used to make the photo. It's the total accumulation of all snow (all photons; all light) that's fallen upon the table that determines how much noise there is and the visibility of that noise in a photo.

As is true of photons emitted or reflected by something, there's randomness in snowfall. While snow might be falling at a different rate of 1-inch per hour in another part of Flagstaff, we've been getting about 2-inches per hour in my neighborhood. That's the rate of snowfall (scene brightness) at my back deck.

In the photo, you'll see the randomness of snowfall exists even at the picnic table level. The surface of the 6 inches of snow atop the table is uneven. That randomness in the depth of the snow, is noise.

You can see the texture in the snow because of variations in its brightness. Those subtle variations add contrast to the scene and make the pebbled texture of the snow visible. That's what noise in a photo looks like. Its randomness in the brightness of a scene that is otherwise evenly lit.

The above photo was made after it had been snowing continuously for about 3 hours. There's approximately 6 inches of snow on the picnic table. The variation in the depth of the snow - the noise - is minor in comparison with the total snowfall - the total signal.

That's how noise works. The greater the total signal (the more snow there is on the table) the smaller the noise portion of that totality will be. Just as randomness in snowfall isn't as obvious when a lot of snow has fallen, noise in a photo isn't as prominent when a lot of light is used to make that image.

If I'd made this photo after it had been snowing for an hour, the depth of the snow per square foot on the table (the exposure) would have been much less. The totality of snow on the table (the total light; the signal) would have been much less. The randomness in snow depth across the table (noise) would also have been less but would have been a greater portion of the total snowfall (signal) at that time. The randomness (noise) would have been more prominent, more obvious to the eye.

This is how noise works in photography. The more light we put on the sensor (or negative, or glass plate) the less prominent noise will be in the photo. This can be achieved by maximizing exposure, which by extension maximizes the total light used to make a photo.

Good explanation; I like it better than the buckets in a field in a rainstorm idea.

And is the bench the smaller sensor?

You took the words right out of my mouth

--
Bill Ferris Photography
Flagstaff, AZ

 

[Quarkcharmed]

The snowflakes work a bit differently than photons. Initially they form random patches (small bumps) but subsequent flakes tend to stick to existing bumps. So it's not a Poisson distribution exactly but maybe close enough as an analogy.

But generally any analogy that follows Poisson distribution would do.

 

[J A C S]

Of course, there is some additional processes, too, but what we see is the absolute noise, not the SNR when we look at the surface.

 

[JustUs7]

And if there’s yellow snow, your white balance is off.

 

[Bill Ferris]

 

[Bill Ferris]

Of course, there is some additional processes, too, but what we see is the absolute noise, not the SNR when we look at the surface.

What I like about the snow analogy is that the variability in the surface of the accumulated snowfall serves as a great visual for noise. There are folks who find it difficult to visualize what noise is. Some think of noise as a particle; kind of like a photon but concealing information rather than adding information.

But noise isn't a particle. It's randomness.

Looking at the accumulated snow, one sees the physical embodiment of randomness. There's naturally-occurring variation that results in snow accumulation that isn't perfectly evenly distributed. That variation and its appearance provides a tangible representation of noise in a photo...as well as exposure and total light.

It may not be as useful for folks who live where snow isn't a regular part of winter. But for the rest of us, it's a useful analog.

 

[Klaus dk]

After a very short exposure, it's even more clear, because the the table isn't completely covered, and the dark tabletop shows through the uneven layer of snow.

 

[Bill Ferris]

After a very short exposure, it's even more clear, because the the table isn't completely covered, and the dark tabletop shows through the uneven layer of snow.

Yes. I think for the visually-oriented person - for instance, a photograher - it's a useful way of seeing the issue.

 

[MJ_Photo38]

This is a good analog, however I'm not sure about the origin of the noise thing.

Yes, there is a random aspect of photons, but that aspect is very much negated by the fact that the density of photons is much much higher than the density of snow flakes falling.

Though visually it can explain noise pretty well on a signal-to-noise-ratio basis, it is not the origin of noise.

The noise we see in our digital images is down to the electronics in the cameras that are more or less able to reject noise. Every component in the system has a noise signature that it puts on the electrical signal that is getting through it.

My point is that noise would rather come from the table than from the snow in this specific case.

a "correct" analogy would be if the table was as rough as the top of the snow (that's the noise level that our sensor has) and then when snow falls onto it, you can observe the patterns on the table get recovered with a pretty good accuracy at first but the more snow falls, the more even-out the table pattern gets and after a few inches of snow you can't see it at all.

Works just as well (even maybe better) with a landscape rather than a table. Without snow (signal) the shapes in the landscape (dirt, grass, rocks that we could call noise) is clearly visible. When it starts snowing, every shape gets a lot smoother and after a while you can't see anything.

Signal is now prevalent over noise, therefore we pretty much only see the signal and the noise is hidden. But if you remove most of the snow (signal) then the rocks, dirt and grass (noise) will start showing up again.

 

[Quarkcharmed]

This is a good analog, however I'm not sure about the origin of the noise thing.

Yes, there is a random aspect of photons, but that aspect is very much negated by the fact that the density of photons is much much higher than the density of snow flakes falling.

That aspect is actually the major contributor to the visible noise.

Though visually it can explain noise pretty well on a signal-to-noise-ratio basis, it is not the origin of noise.

The noise we see in our digital images is down to the electronics in the cameras that are more or less able to reject noise.

Only in very low light/low exposure cases when read noise had a significant contribution. But the photon noise (shot noise) is always there.

a "correct" analogy would be if the table was as rough as the top of the snow (that's the noise level that our sensor has)

A correct analogy is the one that follows Poisson distribution.

 

[Bill Ferris]

This is a good analog, however I'm not sure about the origin of the noise thing.

Yes, there is a random aspect of photons, but that aspect is very much negated by the fact that the density of photons is much much higher than the density of snow flakes falling.

Though visually it can explain noise pretty well on a signal-to-noise-ratio basis, it is not the origin of noise.

The noise we see in our digital images is down to the electronics in the cameras that are more or less able to reject noise. Every component in the system has a noise signature that it puts on the electrical signal that is getting through it.

My point is that noise would rather come from the table than from the snow in this specific case.

The primary source of noise in digital photos is photon noise or shot noise, which is determined by the total light used to make the image. The randomness of the depth of the snow illustrates this. As total depth increases, the ratio of randomness (noise) to total snowfall (signal) decreases. Signal-to-noise ratio (SNR) increases.

Read noise, which you've described, historically gets lower as exposure gets lower. As more manufacturers embrace dual-gain sensor tech and it becomes more common for camera systems to have ISO invariant ranges of performance, the read noise penalty at exposures corresponding to ISOs in the 400 to 3200 range are negated. My Nikon D500, for example, generates as much read noise at ISO 400 as at any higher native ISO.

In very low light situations, read noise - especially pattern noise - can dominate.

But for the vast majority of photography done, shot noise is what we see when looking at photos. That's a big part of what makes the snowfall analog so useful.

a "correct" analogy would be if the table was as rough as the top of the snow (that's the noise level that our sensor has) and then when snow falls onto it, you can observe the patterns on the table get recovered with a pretty good accuracy at first but the more snow falls, the more even-out the table pattern gets and after a few inches of snow you can't see it at all.

Works just as well (even maybe better) with a landscape rather than a table. Without snow (signal) the shapes in the landscape (dirt, grass, rocks that we could call noise) is clearly visible. When it starts snowing, every shape gets a lot smoother and after a while you can't see anything.

Signal is now prevalent over noise, therefore we pretty much only see the signal and the noise is hidden. But if you remove most of the snow (signal) then the rocks, dirt and grass (noise) will start showing up again.

We got another brief snowfall, yesterday. Here's a photo after about 20 minutes of accumulation. The randomness in the depth of the snow is a larger proportion of the total snowfall. SNR is reduced in comparison with the photo I posted the other day. This nicely illustrates the greater impact of shot noise at reduced exposures that deliver less total light to the sensor.

While the table doesn't have the same aspect ratio as a 3:2 digital camera sensor, it better illustrates that element of digital image-making than an open field would. A field is difficult to see or capture in the proper context. It's much easier to visualize the table isolated on the deck as representing a sensor.

--
Bill Ferris Photography
Flagstaff, AZ