NASA's Black Marble team built an algorithm to turn raw satellite nighttime images into usable data about light on the ground. A satellite's day-night band actually measures radiance, a mix of artificial light, reflected moonlight, atmospheric scattering and absorption, cloud cover, snow reflection, and seasonal vegetation changes, not a direct count of electricity use. The Black Marble algorithm estimates and screens out these other contributions, using satellite geometry and additional measurements, so that clouds, snow bands, auroras, and moonlight can be distinguished from real changes in ground lighting.
- 01
The foundational 2018 paper describing this algorithm is credited to first author Roman (name may be misheard in transcript).
- 02
Example images shown include France, where snow in the Pyrenees created a false bright band and clouds obscured cities before correction.
- 03
An aurora appearing in raw satellite data over Scandinavia had to be filtered out as noise rather than treated as real ground light.
- 04
In a US East Coast example (Raleigh, Richmond, Washington DC, Baltimore), an 83% illuminated moon caused clouds to appear falsely bright, which the algorithm had to remove.
1,076 words · auto-generated from the episode video
55:30with paper number one in 2018. This is Rammon. He's the first author. Um Roman, sorry. Roman is the first author. He was um called NASA's Black Marble Nighttime Lights product suite. It's a it's just a product suite that showcases all of the capabilities that this team has created. Ultimately there were sort of especially as a public agency right being backed by NASA they're always trying to create things not necessarily open source might not be the right word but creating platforms and tools that can be utilized >> by whoever >> for endeavors that are fruitful for
56:11their function. Yeah exactly and this is kind of like a methods paper um is what we would call it in the business because this is something that provides a technical foundation for everything else in this episode. So let's start with what exactly a satellite sees. Okay, so it's actually a lot more complicated than you would think. Imagine a patch of ground that's hundreds of meters across. The dayight band in that satellite collects light from the direction of that patch, right? Like my my sensor is pointing over there. It's collecting light from there. Now what arrives in the instrument is something called radiance which roughly speaking is the optical power per unit area per unit angle in that particular range of
56:52wavelengths whatever range of wavelengths the satellite instrument is sensitive to. So it's not a count of illuminated houses. >> It's not an electricity bill for that patch. It includes light that's emitted upwards from artificial sources. That's light that's reflected from the moon. There's light that's reflected through the atmosphere. The atmosphere can absorb light. It's can it can scatter light on its way out. And what the instrument sees depends on the angle relative to the ground as well. >> Right. >> Right. Like if it's looking straight down or if it's looking like sort of towards the curved side of the Earth. And the measured radiance
57:34equals the artificial light that's getting attenuated by the atmosphere plus the reflected moonlight plus all of these other contributions. There's a variety of it. Again, as someone who's not done something in this area, >> in my head I just think, oh, you just point it and everything is perfect. >> Yeah. and I can just have very clear resolution on what I'm looking at. >> Um, but I think as you've eloquently described, >> there are a number of variables between the light that between where the light comes from and it then ultimately being received on the sensor on the satellite. >> Exactly. And so that is where the meat of why these guys got the golden goose
58:16>> Yeah. >> is coming from. It is this. >> It is the black marble algorithm. The job of the algorithm is to estimate and screen those other contributions using the satellit's geometry like where it is in space and where it's pointing and all of these additional measurements like on a cloudy night the ground can be hidden. That observation should be flagged rather than treated as a real power outage. Um on a moonlit night a field may appear brighter that's interesting >> than before. >> Um snow can amplify reflected light of course >> trees can hide light seasonally. That depends on the seasons, right? Because trees in the winter in the in the deciduous forests are not going to highlight as much as So that's not
58:58>> If you were to just not take care of that, you'd be like, "Oh, winter is brighter than summer." No, that's just cuz the trees don't have any leaves, >> right? So there's all of these other >> factors that come into it. And the 2018 paper describes these corrections from the atmosphere, from the terrain, from the vegetation, from the snow, from the stray light, and all of the strategies to fill the gaps in that record. >> This is a a combinatorial problem. >> Exactly. Um, and that's the first major conceptual step. Um, a nighttime photograph, for example, on the left hand side, that's the raw data. >> Mhm. >> And after it runs through the algorithm, we have the stuff on the left. This is in France.
59:39>> Stuff on the on the right. >> Yeah. Yeah. Sorry. The stuff on the right is after the algorithm. The stuff on the left is the raw data that we're getting from the satellite. And you can already see there's so many clouds. >> The Pyrenees have this massive band. That's cuz of the snow there. >> When you take all of that out, then you get the actual cities of France. >> I just can't even imagine. I mean even just in the variables you talk through um because it's not only a function of the different things happening but it's also the different things happening over time. >> Mhm. Yeah. Yeah. It's like a Yeah. It's a time varying phenomenon >> that you have to take out. The noise is extremely dynamic. >> Yeah.
1:00:19>> Right. Um the next one we've got a photo of the north of the Scandinavian provinces. There's an aurora happening right in the raw data. that's not real. You got to take that out, but you got to be careful because there's lights lurking underneath. >> That you'd want to account for. >> And so taking out all of that is the job of the black marble algorithm. >> Unbelievable. >> And the last one, this is for you, Lester. >> Okay. >> Uh this is Raleigh, North Carolina. >> Raleigh. >> Raleigh. Sorry. Yeah, I always mess that up. Um but this is the US East Coast. The moon has a fraction of 83%. So that means like 83% of the moon is
1:01:00illuminated and that's why the the clouds are illuminated in that way. You got to take that out. >> And now you can see Raleigh, Richmond, Virginia, Washington DC, Baltimore is up there. Northern Virginia is also >> this is incredible, >> right? But but but taking all of that data out, that is the job of the black marble algorithm. And that's what the 2018 paper was all about was laying out the foundations of that black marble algorithm. We're taking a raw data set that's not clean >> for this specific use case of identifying
HistotripsySpaceSoft Matter Physics