A Philosophy and Keys to Creating a Thriving Aerial Data Operation — Part 1
In geospatial analytics, aerial imagery data is becoming ubiquitous. There are three main sources of aerial imagery, each which have their own pros and cons and which can most practically be differentiated by collection altitude:
- Low altitude imagery captured by drones. The benefits of drones include the ease of launching and portability, while the downside is the limited coverage you are able to capture due to a low flight altitude.
- Imagery captured by an aircraft, which falls in the low to medium altitude category. This mid-range altitude option will get you a larger swath than a drone, but it will require operating an aircraft, which is generally more costly and labor intensive than a drone.
- Satellite imagery, which is from space. One obvious benefit here is the ability to gather imagery at scale and consistent revisit rates, while obvious drawbacks include sacrificed resolution as well as the sheer cost of launching a satellite — both characteristics which are steadily improving (SpaceX has made a commercial satellite launch available for around a cool $60 million, with analysts projecting that cost to dip around 50%).
All around us, we see examples of imagery from these different sources. Drones are being used for construction surveying (companies like DroneDeploy, Propeller Aero), aircraft are taking images for precision agriculture (Mavrx, Ceres Imaging) or energy infrastructure management (Enview), and satellites are tracking climate data (NASA, NOAA) or retail/logistics volumes (Orbital Insight, Tellus Labs). This proliferation isn’t particularly surprising, as the cost of cameras, launching satellites, and related image processing have fallen dramatically compared to 15 years ago. According to a report by Goldman Sachs from last summer, the market for aerial imagery and related data processing from drones alone is set to hit $100 billion by 2020. That figure doesn’t even include imagery from satellites and airplanes.
Why Is This Important?
There will be opportunities for companies that want to start using the benefits of aerial imagery for business, and a well-run aerial and data processing operation is core to entering this new technology efficiently. However, hundreds of thousands of dollars can be wasted if the proper questions are not asked or answered correctly.
In the following two part post, we will bring up five keys to efficiently running an aerial imagery collection and data processing operation. Part one (this piece) will primarily focus on the data collection aspects of the operation, while the second part will focus on the data processing operation.
1) Understand the Spatial and Spectral Resolution You Need For Your Product
Cameras can cost anywhere from a hundred dollars to many thousands of dollars. A large portion of this cost is the spatial resolution a particular camera will get for you. Typically, this resolution is in part dictated by megapixels, or the number of pixels you will have in an image. Each pixel is the most fundamental unit of analysis for the algorithm that will derive information from the image. Furthermore, the size of the pixel in the real world, also known as the Ground Sample Distance (GSD), can be changed depending on the altitude the image is captured. Hypothetically, if your iPhone 7 with the stock 12.3 megapixel camera was attached to a drone and flown at 1,000 feet, the images would have double the GSD of the same iPhone flown at 500 feet. On the data processing and algorithm side, this means for the given image taken by the iPhone, you will have a total of 12,300,000 color values composing any image, regardless of altitude it was captured at. Depending on different camera parameters, the actual size in the real world which each pixel represents will be different, but in general a lower altitude will translate into a more granular GSD, all other variables held equal.
Another key type of resolution is spectral resolution. While spatial resolution is defined by the number of pixels in a given image and the GSD, spectral resolution is the part of the electromagnetic spectrum that is captured by the camera. For example, humans process images in the “visible”, or roughly the 400 nanometer to 700 nanometer wavelength range of the electromagnetic spectrum, but there are other parts of the spectrum that humans are unable to discern that specific cameras can. A common use case is infrared (IR) cameras in agriculture. For many years, scientific studies have shown certain spectral signatures in the IR correlating to the health of crops. Additionally, having the right electromagnetic width and number of narrow bands is critical in value-added agricultural markets, such as nitrogen management or water stress measurement.
The key question to ask here is: Is the GSD and spectral range acquired good enough to extract something valuable out of the image? If a company wants to create a product that surveys an above-ground natural gas pipeline that is 20 cm in diameter, will a pixel that represents 40 cm in the real world accurately capture where the pipeline is? On the flip side, is a 40 cm GSD absolutely necessary to discern urban development change, vegetation cover change, or measurement of coastline erosion? If the GSD was one meter instead, we most likely would still be able to quantify change fairly accurately while driving down costs.
2) Understand the Frequency of Capture Needed for a Compelling Product
Significant change in a city is slow, construction sites and agriculture plots can evolve every few days, and a disaster can change by the minute. The more frequently imagery is captured, a more complete story can be told, however, the cost of acquiring the imagery also increases. A key concept in setting the frequency of capture is understanding the opportunity cost of not capturing data. For example, a drone or aircraft can easily capture 100,000 km of oil or gas transmission pipes over one year, but can a leak go undetected for that long? How much will the operating expenditure for the oil or gas company increase in the time that leak is undetected because each segment of the network was surveyed only once a year?
On the contrary, the benefits of capturing a small city with little visible aerial change on a frequent basis has little value for an imagery provider or user, as the user will most likely not gain anything from this new capture.
The key question to ask here is: If the data was not captured, how would this lack of data affect my end product? Taking agriculture as a hypothetical example, would a farmer be able to go without actionable data for one day, three days, or a week? At the end of the day, the frequency should be dictated by the value added from each capture compared to the cost incurred through the capture.
In this post, we focused on the data collection aspect of the aerial operation. Key factors to explore include:
- Setting the spatial and spectral resolution. Megapixels, GSD, and the electromagnetic range collected are all part of the equation.
- Dictating the frequency of capture. This should be thought about in the context of opportunity cost (what is the cost of not capturing data?) and a frequency timetable should be built off of that analysis.
While the above points are important, an image by itself is not of significant use — it needs to be processed and integrated into a product that will add value to an industry customer. Typically, the data will need to be uploaded to some sort of processing environment, where algorithms can be applied to the imagery.
Part two of this post will focus on key questions to explore in the data processing side of the operation.
Alim can be reached via message on LinkedIn. He is open to discussing the subject matter and exploring new opportunities in the industry. Alim previously led 3D Aerial Operations at Google, however, the views of this article only reflect Alim’s personal philosophy on the topic.