A Unique Study: Identifying Flying Airplanes Over a Very Large Region Using 10m-Resolution Sentinel-2 Satellite Imagery
Important airport features, such as airplanes, runways and terminals (buildings), are routinely monitored by high-resolution aerial and satellite imagery. But flying airplanes can only be observed occasionally due to the relatively small swath associated with each high-resolution image scene.
This post discusses the potential of medium-resolution satellite imagery (10m-resolution Sentinel-2) for identifying many flying airplanes over a very large region.
Spatial Resolution and Airplane Identification
The dimension of some most common passenger airplanes is shown here:
- Airbus A320: 37.57 m(length) x 35.8 m(wingspan)
- Boeing 747–400: 70.6 m (length) x 64.4 m (wingspan)
Obviously, low-resolution imagery cannot capture airplanes of these sizes and a finer spatial resolution is a must. But, which resolution level would be the limit for the airplane identification in imagery? Let’s take an example first, and compare the 10m-resolution Sentinel-2 imagery with the 15m-resolution Landsat-8 imagery (both are freely available), for the London Heathrow Airport (Figure 1):
Through this comparison in Figure 1, it is fair to say that the 15m-resolution Landsat-8 imagery is over stretched for identifying airplanes, whereas the 10m-resolution Sentinel-2 imagery can still record sharp and clear shapes of common passenger airplanes resting on airport fields. If the size of an airplane (e.g. small private jets) is too small, obviously the Sentinel-2 imagery would not be useful. The spatial resolution of the imagery is the first most important factor for identifying airplanes, be they resting or flying.
Identifying Flying Airplanes from Sentinel-2 Imagery
Modern satellite sensors, such as Sentinel-2 multi-spectral instrument (MSI), often use a push-broom design, and the imagery for each spectral band (e.g. blue, green, and red) is captured with a very minor delay in time. For fast moving objects on the ground (e.g. cars on freeway) or in the air (e.g. airplanes), the pixel position of the same object is actually slightly shifted from one band to another. Figure 2 shows such an example with three airplanes identified each in a different moving stage (still from the same Sentinel-2 imagery for the London Heathrow Airport):
- Taking off
This is a generic feature for the new Sentinel-2 imagery, but not the Landsat-8 imagery. We explore another Sentinel-2 imagery tile for more vivid examples (Figure 3). Given the identified flying trajectories, relative pixel shifts of flying airplanes across bands (in this case, red/green/blue natural color bands), and the known sensor metadata (such as height, viewing incidence angle, travelling speed and acquisition time gaps across bands), one may be able able to estimate the speed of flying airplanes.
A few more examples from other international airports are shown below:
- Paris: Charles de Gaulle Airport (Figure 4)
- San Francisco: San Francisco International Airport (Figure 5)
- Osaka: Kansai-Airport Station (Figure 6)
Applications: Identifying Many Airplanes Over a Very Large Region Using 10m-Resolution Sentinel-2 Imagery
Sentinel-2 satellite has a very wide swath (290 km) and a high frequent revisit interval (5 days at the Equator once the twin satellite is to be lunched and operational in 2016/2017). These features, along with the fine spatial resolution of 10 m (for three visible bands and a near infrared band), enable its potential application for identifying airplanes over a very large region. Figure 7 compares swath width among Sentinel-2, Landsat-8, and WorldView-3; in terms of scene sizes, a Sentinel-2 scene can cover ~490 times more areas than that of a very high-resolution WorldView-3 scene.
Taking a recent Sentinel-2 tile (about 100 km by 100 km in size) near the Istanbul Ataturk Airport as an example, we can easily identify many flying airplanes (Figure 8). It is done by manual interpretations, a method that is acceptable for this type of special applications. The classified result might be useful for documenting and investigating air traffic, and complementary to near real-time air traffic mapping (e.g. Flightradar24).
(A side comment: For automated feature classification from such imagery sources, modern methods such as artificial intelligence and machine learning are being actively explored these days, but their classification effectiveness and accuracy still need to be more openly and critically evaluated case by case by professional peers.)
A zoomed view is shown in Figure 9, and its full resolution (10 m) picture can be downloaded here.