In a Q&A with Dedrone’s Director of Engineering Michael Dyballa, we break down the connections between neural networks, machine learning, and deep learning to best understand how Dedrone’s proprietary DroneDNA is the only database of its kind, and how it advances every single day with all drones we detect.

What is DroneDNA?

DroneDNA is Dedrone’s database of drones and other aerial objects, such as birds, helicopters, and planes. DroneDNA is a critical element to our DroneTracker system, which combines information from various sensors, such as radio frequency and Wi-Fi scanners, to determine if an object in the sky is a drone.

How is the DroneDNA library built?

We procure information through our own and our customer’s cameras around the world, and map, or annotate, the elements. Once the images are annotated, we load them to the library. This information is analyzed by the DroneDNA process, which learns from the images to make a positive identification of a drone.

Additionally, as every new drone is manufactured and available on the market, we break it down, learn from it, and add visual data and characteristics to the DroneDNA. This process happens within a matter of days, and we regularly add updates to the network.  

With every new drone and customer, we’re learning more about our environment. It’s a dynamic process.

How can you teach software that an image is a drone and not another flying object?

There are layers of information in every single image. The size, shape, structure, shading and then how an object changes its perspective when it moves even just a centimeter in any direction. A drone has very specific characteristics, and we’re annotating all flying objects detected through cameras, such as like birds, helicopters, and passenger aircraft.

Manual annotation is completed by the Dedrone software team to create this precise database. These images are uploaded to the DroneDNA database, and from there, the database references this new material, learns and advances from the previous injection of information. This is the basis of machine learning, which is a very popular intersection of computer science and mathematics - it’s the art of teaching computers and software based on data.  

How big is DroneDNA?

Each time we update DroneDNA, we process over 250 million different images of drones, aircraft, birds and other objects. In the past 8 months, we’ve annotated 3 million drone images.The number of combinations of visual information is colossal. As we add more images, the database becomes deeper and more sophisticated. This is the basis for deep learning, as DroneDNA uses layers of information to make decisions.

How does the DroneDNA database learn from images?

DroneDNA is a machine learning program and was built based on a neural network, which is a system initially defined by neuroscientists to map out the brain and see how we make decisions.

Neural networks have been popping up in recent years as technology engineers are developing sophisticated programs which learn from the information that is input into software, and make decisions. This type of network is used in developing artificial intelligence, driverless cars, and precision medicine and now, for detecting drones.

Similar to how we process something through our eyes, when we see a drone, we see more than just the physical hardware. We look at “visual signatures,” including the shape, movement, and shadows. Each of these elements is processed through our brains to decide whether or not something is a drone/bird/plane. DroneDNA has the same capability.

Each image we add to DroneDNA adds to this machine learning and decision-making process. DroneDNA answers two simple questions: “Is this a drone and what kind of drone is this?” Each element we annotate is used to get a more precise response. Every time we detect a new drone, DroneDNA learns from those incidences to improve the performance of our systems. Since we extract patterns of many different drones, our system is able to detect even unknown/DIY drones.

Do you just need a camera to decide if an object in the sky is a drone?

DroneTracker can detect drones over a mile away from the protected site, which is sometimes beyond what a camera can clearly see. This is why DroneTracker connects to multiple sensors, like Wi-Fi and radio frequency and microphones. The fusion of this information, along with the camera and DroneDNA data, allows for supreme detection capabilities and very quick reaction times to deploy security measures.

What's next for Dedrone and DroneDNA?

We’re regularly adding images to the database. While it’s a large network now, it’s getting bigger each day. As the drone market expands, we’re seeing new drone models sold, and watching them be used in innovative ways for different industries and environments. Every new application for a drone is helpful for us to learn from, which is the core of our work with DroneDNA - it’s constantly being updated and evolving to provide the most accurate detection process in the world.

Susan Friedberg

About the author

Susan is the Director of Communications at Dedrone and has researched, developed, and led the conversation on drones, counter-drone technology, and airspace security since 2016.

Susan Friedberg

Originally published May 3, 2017, updated Dec 4, 2023