Over the past few decades, the availability of commercial satellite images has dramatically changed how academics and other nongovernmental analysts study the spread of nuclear weapons and ballistic missiles. Most analysts have made use of high-resolution images comparable to the pictures taken by spy satellites during the Cold War.
Sharp, high-resolution images offer obvious advantages to the analyst, but they also have certain limitations. For example, there is a direct tradeoff between how sharp a picture is and how large of an area can be photographed. High-resolution images have a narrow field of view.
This introduces a number of complications. The relatively small footprint means that high-resolution images remain scarce, with high-priority areas attracting the most coverage. But it is not always obvious where analysts should be looking. Nuclear weapons and missile programs are often secret. That means activities are likely to occur outside known locations, either at purely secret locations or locations previously considered to be of low interest. Looking at out-of-the-way locations can be difficult—it can be cost-prohibitive to task a satellite, and the analyst may worry about investing scarce resources on locations that don’t pan out.
At the same time, the commercial space sector is developing many constellations that are outside of the traditional high-resolution optical image. These include large constellations of moderate-resolution imaging satellites that provide near-daily coverage of the earth and multispectral and radar capabilities with regular and predictable revisit times. What connects these “new space” constellations is the importance of using software, such as ENVI from L3Harris Geospatial, to process the images.
For analysts studying the spread of nuclear weapons and ballistic missiles, searching for new or unknown facilities is a significant challenge. Searching high-resolution images is cost-prohibitive and inefficient. At the James Martin Center for Nonproliferation Studies (CNS), we are increasingly using high-cadence moderate-resolution (3 m) images provided by Planet and even low-resolution (30 m) data available from LANDSAT.
Interpreting such imagery, however, requires software and the training to use it. While interpreting high-resolution images is analytically demanding and aided by software, the task still involves an analyst looking at an image. However, at lower resolutions, it can be much harder to identify and interpret changes with the human eye. Analysts working with these images must rely on software, such as ENVI, to present the data in the image in usable form.
For example, consider these two LANDSAT images of a site near Dawadmi in Saudi Arabia. If one looks extremely closely, some differences may appear. But with only 30 m resolution, this is some of the hardest imagery to interpret. Even to a trained eye, any differences are relatively small.
Left: An April 2013 image of an area near Dawadmi, Saudi Arabia. Right: A November 2014 image of the same location. (Landsat-8 images courtesy of the U.S. Geological Survey)
Using high-resolution images, an analyst writing in Janes’s in 2013 first identified this site as a missle base in Saudi Arabia. Analysts at CNS were aware of the identification but did not accord it a very high priority. Neither did others. We checked it every few years when high-resolution satellite imagery became available for free on Google Earth. Had we been able to fully exploit low- and moderate-resolution satellite images, we would have noticed dramatic changes occurring within a few months of the site’s identification—changes that are visible in those two LANDSAT images.
Looking for change at similar locations using low-resolution imagery such as this LANDSAT image requires thinking analytically about land use at the site. While there is much discussion about automated change detection using machine learning, the task is not as simple as asking a computer to compare two images of the same location. For example, here is the same pair of LANDSAT images analyzed with a commercial change-detection algorithm. Notice the algorithm picks up the changes in darkness in shadows and nothing else. There is nothing wrong with the algorithm—it is doing exactly what it has been asked to do. But we’ve asked the wrong question.
Above: A commercially available change detection algorithm applied to the same LANDSAT images. (Landsat-8 images courtesy of the U.S. Geological Survey)
A better approach is using spectral classification to determine land use. The idea is to slice the image into many spectral classes because one land-use application, like farming, will have a different spectral signature than a different use, such as a building. This is the same approach that companies use for road or building extraction from satellite imagery, as well as resource exploration.
While conceptually straightforward, isolating the spectral signature for different land-use applications can be extremely hard because they vary tremendously by locality. This is a problem for nonproliferation analysts because we are often looking at unusual or remote places in the world where there is not a well-developed spectral library for land classification.
ENVI, however, offers a powerful tool that allows the user to produce a bespoke classification, either in the form of a “supervised” classification or an “unsupervised” classification. Supervised classification involves training an algorithm on the spectral signatures of a specific location by teaching it to recognize buildings, roads, and agricultural activities and categorize pixels into different classes.
ENVI also offers the option to choose “unsupervised classification,” in which the software breaks the image into an arbitrary number of spectral classes selected by the user. The latter approach is extremely flexible, especially for an analyst working with landscapes that do not have well-defined spectral libraries for local land use. In many cases, our goal is to quickly identify a spectral signature that is good enough to reveal changes that interest us—such as the construction of new buildings. In this case, we broke the image into 30 classes. (See the image below, left).
Left: The same LANDSAT-8 image of Dawadmi divided into 30 spectral classes using the ENVI classification tool. Right: The same image with just one of the classes – Class 30 – isolated.
The idea is to isolate any spectral classes corresponding to human development—buildings, roads, cleared earth, etc. We quickly found that most buildings were captured by Class 30 (displayed in red above). While these buildings are difficult to make out with the human eye in the 30 m LANDSAT image, the algorithm does an excellent job of picking them out, as can be seen in this overlay with a high-resolution image from the same time period.
Above: The LANDSAT-8 image with class 30 overlaid on a high-resolution image in Google Earth. (Source: Google Earth/Maxar)
Creating a bespoke spectral index for land use in a particular location is typically a significant research project. As nonproliferation analysts, however, we’re interested in a different question—and the land-use classification provided in a few seconds by ENVI is more than good enough for our use case. In the example above, Class 30 captures both buildings and some disturbed earth. It does miss some human-made features, but overall, it captures most of what we’re interested in finding.
Having established a specific land-use signature from the 2013 image, we can then perform the same analysis on an image taken about 18 months later. We can see a substantial increase in areas where there are either buildings or disturbed earth.
Left: Class 30 in the April 2013 LANDSAT-8 image. Right: Class 30 in the 2014 LANDSAT-8 image. (Landsat-8 images courtesy of the U.S. Geological Survey)
The second image indicates that there have been substantial changes to the site. As it turns out, the classification worked exceptionally well. When we overlay the second image with a moderate-resolution satellite image taken by Planet, we can see that the changing land use detected by the ENVI classification tool corresponds to a significant expansion of the facility. The correspondence is not perfect, but it does not need to be. Our goal is for the software to tell us whether the change justifies a closer look or not.
Above: Class 30 from the April 2013 LANDSAT-8 image overlaid on an image from Planet Labs. (Landsat-8 image courtesy of the U.S. Geological Survey. Planet Labs image cc-by-nc-sa 4.0.)
Although the construction was visible in late 2014 in LANDSAT data, we did not notice the change until 2018. Even so, we were the first organization to take note of the new construction and worked with Planet to task a high-resolution Skysat satellite to take an image of the facility. We were then able to assess that this site was a Chinese-supplied facility to produce ballistic missiles. While there is no international legal prohibition on such activity, it was still a significant policy development in a region where many experts are worried about a burgeoning race in long-range ballistic missiles. Members of Congress took note of the development and pursued it. Several months later, in late 2019, U.S. officials confirmed that the United States intelligence community had reached the same conclusion.
We did not, however, need to wait until 2018. The scale of construction was visible in 30 m LANDSAT images by late 2014, had we used the correct software and employed the correct analytical approach. Moreover, the changes in land use were noticeable enough that we could have automated the workflow, pulling new images when they are available and notifying analysts of any significant change in land use. Had we done this, we could have tasked a satellite several years earlier before the facility was completed, which might have had a significant change in the outcome of the policy debate.
So, while most of the attention in the media goes to high-resolution images that are easy to understand, analysts have to use the full spectrum of data available from space. That spectrum extends to include new capabilities such as high-cadence and multispectral imaging where resolution is less important. These new constellations offer customers tremendous value—but properly exploiting them requires software like ENVI. Commercial firms increasingly see that value. As a result, these new constellations are reshaping the so-called “new space” sector. They can also reshape how we study the spread of nuclear weapons—but only if we invest in a generation of analysts capable of using software to fully surface the value in them.