Партнерство Planet и Metaspectral запустит анализ гиперспектральных данных для мониторинга экологии и климата

I think it’s fair to say that I’ve been blessed with a cheerful and optimistic disposition. I fully relate to the classic Always Look on the Bright Side of Life song from Monty Python’s Life of Brian movie. If you ever wondered about the origin of this song, I recommend Always Look on the Bright Side of Life: A Sortabiography by Eric Idle, but I fear we’ve already meandered into our first digression. I don’t wish to boast, but wandering off into the weeds this early in a column is a bit of a record, even for me. The thing is, I’m growing increasingly despondent with respect to our collective future as a species. I want us first to become an interplanetary civilization and, later, an interstellar civilization. However, I worry that before we achieve either of these dreams, we may severely degrade Earth’s ability to sustain our existing civilization. On the bright side, as discussed in Alan Weisman’s The World Without Us, Earth itself will be fine in the long term. On the downside, it’s our ability to live pleasantly on it that’s at stake. I’ve read a lot of postapocalyptic science fiction books. Many of these are based on scenarios where we devastate the world by means of nuclear warfare. I’m thinking of tales like Songs from the Stars by Norman Spinrad, The Chrysalids by John Wyndham, and Riddley Walker by Russell Hoban. Others are based on the effects of global pandemics, like Eternity Road by Jack McDevitt, The White Plague by Frank Herbert, and The Stand by Stephen King. But the ones we’re interested in here involve catastrophic climate-driven disasters, like The Drowned World by J. G. Ballard, The Water Knife by Paolo Bacigalupi, and New York 2140 by Kim Stanley Robinson. The disturbing thing is that some of these fictional futures are starting to feel a little less fictional. In Paolo Bacigalupi’s The Water Knife, for example, the American Southwest is ravaged by drought as states and cities fight over the dwindling waters of the Colorado River. Back here in 2026, federal officials have just announced additional water cuts as the real Colorado River continues its decades-long decline. Meanwhile, on the other side of the Atlantic, record-low water levels on the Danube have exposed long-submerged World War II warships and forced a Bulgarian nuclear power station to reduce its output. The Rhine has been running so low that freight barges are being forced to operate with substantially reduced loads. And, while all this is going on, sea levels continue their inexorable rise, making Kim Stanley Robinson’s vision of a partially submerged Manhattan in New York 2140 seem rather less fanciful than one might wish. I don’t wish to be alarming, but… (Source: me and ChatGPT) This leads us to Sarah Wilson’s recently published I Eat the Stars: How to Live Fully and Beautifully in a Collapsing World. Sarah’s thesis is considerably more disturbing than its rather whimsical title might suggest. She argues that climate change is only one component of a much larger problem: modern civilization has become an enormously complex system of interconnected systems, increasingly stressed by climate change, resource depletion, inequality, political polarization, conflict, and other pressures. Her conclusion is that some degree of systemic collapse may now be unavoidable (I bet she’s a wow at parties). On the other hand, knowledge is power, and the more we understand what’s going on, the more we can do about it. All of which leads us to the reason for today’s meandering musings, which is that I recently got to chat with Francis Doumet, who is the CEO of Metaspectral. When I first wrote about Metaspectral three years ago (see Metaspectral Takes AI Processing of Hyperspectral Imaging by Storm), much of our conversation involved hyperspectral cameras close to the ground. One example that particularly stuck in my mind involved recycling facilities, where visually indistinguishable pieces of plastic could be racing along conveyor belts. They might all look like transparent plastic to you and me, but hyperspectral imaging could reveal that they were made from completely different materials, allowing them to be identified and sorted accordingly. Metaspectral’s technology has since found its way into a range of applications, including aerospace, defense, agriculture, mining, environmental monitoring, and manufacturing. This time, however, we’re going to raise our sights… all the way into space. Before we get there, however, it would probably behoove us to make sure we’re all tap-dancing to the same drumbeat regarding spectral imaging. Consider an ordinary digital camera. At the risk of oversimplifying horribly, we can think of it as measuring three broad bands of the visible spectrum: red, green, and blue (RGB). Measuring the relative amounts of light in these three bands at every pixel is enough to create the full-color images we know and love. The next step up is multispectral imaging. Instead of measuring only red, green, and blue, a multispectral sensor might use five, six, eight, or a dozen bands, some of which may extend beyond visible light into regions such as the near infrared. Even this small increase in the number of bands can reveal things that aren’t apparent to the naked eye. Hyperspectral imaging takes this idea and runs with it. Rather than a handful of relatively broad bands, the spectrum is divided into hundreds of narrow slices. Take the company Planet, for example. Planet’s Tanager hyperspectral satellites capture 424 spectral bands spanning wavelengths from 400 to 2,500 nanometers. In other words, they start in the visible portion of the spectrum and continue far beyond anything our eyes can perceive into the shortwave infrared. Why go to all this trouble? Well, suppose we’re looking at two patches of vegetation. To our eyes, and perhaps even to an ordinary RGB camera, they may appear to be essentially the same shade of green. But one patch might be healthy while the other is suffering from water stress, disease, or a nutrient deficiency. These differences can lead to subtle changes in how the plants absorb and reflect specific wavelengths of light, enabling hyperspectral imaging to detect issues before visible symptoms appear. Indeed, Metaspectral says this is one of the potential agricultural applications of its platform. The same idea applies to minerals. Two rocks may look much the same shade of brown or gray to us, while exhibiting quite different spectral characteristics. Different plastics can look identical in visible light while being distinguishable spectrally. Soil, water, vegetation, minerals, man-made materials, and other substances all interact with different wavelengths of electromagnetic radiation in their own characteristic ways. Another way of thinking about this is that every pixel in a hyperspectral image comes equipped with its own spectrum. Instead of associating three RGB values with a pixel, we might associate hundreds of measurements representing the amount of light detected in hundreds of narrow wavelength bands. Plot wavelength along the horizontal axis and the intensity of the reflected light along the vertical axis, and we end up with what is known as a spectral signature. Francis says Metaspectral has a library containing the spectral signatures of tens of thousands of compounds. During our conversation, we came up with another way to visualize all of this. Imagine the image we see on our screen as a single playing card. Now imagine a hyperspectral image as a whopping great pack of hundreds of cards, all showing precisely the same scene, but with each card representing the intensity of the light reflected in one narrow spectral band. That’s not technically how the data has to be stored, of course, but it’s a useful way of wrapping one’s brain around the concept. All of which is wonderful, except that we’ve just created another problem. We’ve gone from three measurements per pixel to hundreds. Even a relatively modest image containing one million pixels can therefore contain hundreds of millions of spectral measurements. Francis says a single hyperspectral scene can easily occupy around a gigabyte, while systems continuously capturing this sort of data can generate tens of terabytes per day. Furthermore, there’s a big difference between possessing all this data and knowing what it means. With an RGB photograph, we humans can look at the picture and make sense of it. If a river has burst its banks and flooded the surrounding countryside, for example, there’s a good chance we’ll be able to see this for ourselves. But none of us can look at hundreds of spectral bands—many of which lie outside the range of human vision—and intuitively determine what they are telling us. Historically, this sort of analysis required specialists in remote sensing, perhaps supplemented by domain experts such as geologists or agricultural scientists. And this, of course, is where artificial intelligence starts rubbing its metaphorical hands together with glee. One of AI’s great strengths is its ability to sift through enormous quantities of data, discover patterns and correlations that would overwhelm a human, and determine which pieces of information are significant for the problem at hand. Rather than requiring a human expert to say, “If you’re looking for this mineral, concentrate on bands 37, 96, and 217,” we can provide the AI with the entire hyperspectral dataset and let it determine which spectral features matter. Metaspectral has spent years developing specialized deep-learning models for precisely this purpose. These models are designed to tease out rare or hidden signals buried in noisy hyperspectral data. Metaspectral’s tools perform functions such as spectral unmixing, which can tease apart the different materials contributing to a single pixel; target detection, which searches an image for the spectral fingerprints of particular materials or substances; and regression, which uses spectral information to estimate quantities such as material concentrations or properties. Francis likens some of these applications to finding a needle in a haystack, and says these specialized models help increase accuracy while reducing false positives and false negatives. This brings us to Metaspectral Clarity and its newer agentic AI capabilities. An important point is that the agentic AI isn’t being asked to perform all of the hyperspectral analysis itself. Metaspectral already has its own specialized spectral-processing and deep-learning tools. Instead, the agentic AI acts more like the conductor of an orchestra, determining which tools to call, in what sequence, and what to do with their results. This means a user doesn’t necessarily need to be a remote-sensing scientist. Francis offered the example of someone interested in mineral exploration identifying an area of interest and essentially asking, “Where should I look for copper in this area?” The AI agent can determine which minerals may indicate the presence of copper in that particular geological setting, find the appropriate reference spectral signatures, apply Metaspectral’s analytical tools to the hyperspectral data, and present its findings to the user. There’s another aspect of all this that’s especially important when the results are used to make potentially consequential decisions. “Because the AI said so” isn’t a particularly satisfying answer if we’re talking about agriculture, mining, natural resources, environmental monitoring, or defense. Metaspectral’s goal with Clarity is therefore to provide not just an answer, but the evidence underlying that answer, including the spectral evidence, confidence metrics, explanations of how the conclusion was reached, and—where appropriate—supporting scientific literature. All of which brings us back to Planet, whose rather wonderful mission is: “To use space to help life on Earth.” Planet operates a large fleet of Earth-observation satellites, and its Tanager mission adds hyperspectral imaging to the mix. This represents an extraordinary source of information about what’s happening on the surface of our increasingly beleaguered little blue-green world. The problem, once again, is that collecting information and understanding it are two very different things. This is where the recently announced partnership between Planet and Metaspectral comes into play. The division of labor is easy to understand: Planet provides the imagery; Metaspectral provides the tools to analyze it. The companies themselves express the idea slightly differently. From their perspective, Tanager provides the hyperspectral observation layer, while Clarity provides the decision-intelligence layer. The goal is to transform vast collections of orbital hyperspectral measurements into useful, reviewable, explainable information for applications such as agriculture, forestry, environmental monitoring, mineral exploration, and change detection. And this is where we circle back to my earlier musings about climate change, ecological degradation, and the long-term prospects for our civilization. If we’re going to look after our planet, one of the first things we need is the ability to understand what’s happening to it. Planet gives us an extraordinarily powerful new set of eyes in space. Metaspectral gives us the intelligence to understand what those eyes are seeing. Which means that, for once, what may initially have appeared to be one of my rambling digressions turns out to have been relevant all along. Who would have thunk it?