Identifying aircraft types with an app: AI image recognition and its limits

You are standing at the fence, a narrowbody jet rolls by, and you are not sure: A320 or A321? neo or ceo? The temptation is big to just throw the photo into an app and let it read off the type. Identifying an aircraft type with an app sounds like the perfect job for artificial intelligence – and partly it is. This article shows you how AI image recognition works on aircraft, where it is reliable, where it systematically misses and how to back up its suggestions so that the right type ends up in your logbook.

The short answer

AI image recognition can reliably determine the rough aircraft class and many distinctive types: four-engine jets, widebodies with an upper deck, regional turboprops or Airbus narrowbodies. As soon as it comes to closely related variants – fuselage stretches, engine generations, subtypes – the hit rate drops noticeably. For a clean logbook, it is therefore a good first suggestion but no substitute for the registration and your own look at the details.

It is also important that not every app that “identifies aircraft” looks at the image to do so. Many spotters confuse three fundamentally different methods, each with its own strengths and weaknesses.

Three ways apps “identify” aircraft

1. True image recognition

Here a neural network analyzes the photo itself. The model was previously trained on a great many labeled aircraft images and has learned which image patterns are typical of which type: the ratio of fuselage length to wingspan, the shape of the nose, the number and position of the engines, the tail. General image searches like Google Lens and multimodal AI assistants to which you send a photo belong in this category. They usually deliver a type suggestion, sometimes with a rationale, often without saying how sure they are.

2. AR view with position data

Flight tracker apps like Flightradar24 or Plane Finder offer views in which you point your smartphone at the sky and the app labels the flights in the camera image. That looks like image recognition but is not: the app knows your position via GPS, your viewing direction via compass and orientation sensor and overlays which flights, according to live data, are currently in that direction. It takes the type from the database that belongs to the aircraft’s transponder code. How this data comes about is explained in the guide Flight tracking with ADS-B for spotters; which services offer which views is covered in the flight tracking apps comparison.

3. Registration lookup

The third method uses what is on the aircraft anyway: the registration. Whether you read it yourself or text recognition pulls it from the photo – with the registration you look up in a fleet database exactly which subtype is behind it. How registrations are structured is explained in Understanding aircraft registrations, and which databases are worthwhile in the fleet databases comparison.

Method	What it evaluates	Strong at	Weak at
Image recognition	Shape and details in the photo	distinctive types, rough classes, photos from the archive	closely related variants, small or blurry subjects
AR with position data	GPS, compass, live flight data	aircraft high in the sky, quick live identification	aircraft without received data, dense traffic, inaccurate sensors
Registration lookup	Registration plus database	exact subtype, year of manufacture, history	distant aircraft, hidden or illegible registration

How AI image recognition works on aircraft

At its core, image recognition is classification: the model gets an image and outputs a probability for each class it knows. The class with the highest value is shown as the result. The model does not “know” what a sharklet or an engine pylon is – it has only learned which pixel patterns appeared together with which label in the training images.

In research, aircraft are considered a classic example of fine-grained classification. This refers to tasks in which the classes look very similar and differ only in small details – similar to bird species or car models. A well-known test dataset for this is FGVC-Aircraft from the University of Oxford from 2013: around 10,000 photos, mostly from plane spotters, sorted into just over 100 model variants, families and manufacturers. That researchers chose aircraft in particular as a especially difficult test task says a lot about why apps hit limits here.

Three characteristics follow from this that you should know:

• The model only knows what it has learned. A type that is missing from the training data will inevitably be recognized as something else – usually as the most similar known type. Rare types, vintage aircraft or military variants are then not hit.
• Common types are preferred. Types that appeared often in the training images win when in doubt. A rare subtype is therefore readily sorted into its widespread relative.
• Liveries can deceive. If a model learns that a certain airline livery was almost always seen on a certain type, it can weight the color more heavily than the shape. If the same airline flies several types, it becomes uncertain.

Where the AI is reliable

It would be unfair to talk image recognition down. For some tasks it is genuinely helpful:

• Rough classification: Twinjet or four-engine jet, jet or turboprop, narrowbody or widebody, helicopter or fixed-wing aircraft – good models usually tell these apart reliably.
• Distinctive silhouettes: An Airbus A380 with its full-length upper deck, a jumbo with its hump or a type with a very distinctive shape is easy to grasp even for a model. The difference between the two big double-deck types is explained in Telling the A380 and Boeing 747 apart.
• Manufacturer level: Whether a narrowbody jet comes from Airbus or Boeing is often correctly identified with a good view of the nose and cockpit windows. The features that help with this are described in Telling Airbus and Boeing apart.
• Archive work: If you have hundreds of old photos without labels, an AI delivers a quick first suggestion that you then only have to check. How to store the results permanently in the image is shown in Labeling photos properly.

Where the AI regularly misses

The weaknesses concentrate on exactly the cases that spotters care about most: subtypes and variants. The following overview shows typical problem pairs and the features you should look at yourself.

Confusion pair	Why the AI is uncertain	What you look for
A320 and A321	same nose, same wings, only the fuselage length differs	fuselage length relative to wingspan, number and position of the emergency exits over the wing
A320ceo and A320neo	identical overall look, the difference lies in the engines	engine diameter and nacelle shape, see A320neo or ceo
Boeing 737-800 and 737 MAX 8	almost the same length and silhouette	tail cone, serrated engine nozzles, winglet shape, see Boeing 737 generations
787-8, 787-9, 787-10	same family, only the fuselage length differs	ratio of fuselage length to wingspan, distance between wing and rear doors
A350 and Boeing 787	similar era, similar proportions from a distance	cockpit window mask, wingtips, nose, see A350 or 787
E-Jets and CRJ	similar size, both regional jets	engines on the wing or at the tail, see Telling regional jets apart

With the A320 family, the problem is especially clear. Between the A320, A320neo, A321 and A321neo, a model often sees only a few pixels of difference if the aircraft does not fill the frame. A look at the fuselage length and the engines is more reliable here – the article Telling the A318, A319, A320 and A321 apart shows what matters.

Similarly with Boeing: the 737-800 and 737 MAX 8 have almost the same silhouette from a medium distance. Once you know the differences, you see them faster than any app.

What your photo is worth for identification

Whether image recognition hits depends heavily on the source material. A model can only use what is visible in the image. Typical interference factors:

• Small subject: A jet at cruising altitude or far back on the apron consists of only a few pixels. Engine shape, doors and wingtips are lost.
• Backlight: As a dark silhouette, all details on the side of the fuselage are missing. More on planning for light is in Planning sun position.
• Heat shimmer and haze: Over warm asphalt, fine edges blur. For the AI, this looks similar to blur.
• Cropped subjects: If the nose or tail is missing, often exactly the features that separate two variants are missing.
• Unusual perspectives: Photographed from directly below, head-on or from diagonally behind, aircraft look different than in most training images, which tend to show side views.
• Night: In darkness, often only lights remain. How to classify aircraft anyway is explained in Identifying aircraft at night.

Conversely: a sharp, frame-filling side view in good light is the ideal material – for the AI as for you. Even then, it is worth reading the registration too, because it is almost always clearly legible on the fuselage in such images.

AR view: accurate in the sky, with its own pitfalls

For overflights at high altitude, an AR view in a flight tracker app is usually the better choice than any image analysis, because it does not have to “see” the aircraft at all. But it has its own sources of error:

• Sensor deviations: A poorly calibrated compass shifts the overlays. With several aircraft close together, the label can land on the wrong object. Recalibrating helps, and so does cross-checking in the info panel.
• Missing data: Aircraft whose signals are not received in your area or that are filtered by services do not show up. This happens especially with military aircraft and some business jets.
• Database errors: The type shown comes from a mapping between transponder address and aircraft. If it is outdated, for example after a change of owner, the type information is wrong too. Background on this is explained in ICAO 24-bit address (hex code).

On the ground, on the apron or near the runway, the AR view helps less because many aircraft stand close together there. Here, looking directly at the registration and features is usually faster.

How to verify an AI suggestion in five steps

The best way to handle image recognition is a fixed verification routine. It costs hardly any time and keeps errors from piling up in your logbook.

Which route for which situation?

Not every method fits everywhere. The decision depends on where the aircraft is and what you know about it.

If you work with registrations, you can also search specifically for individual aircraft. How to find and track a particular aircraft via its registration is shown in Finding an aircraft by registration.

Privacy and rights when uploading

Many AI services do not process photos on your device but on their own servers. Before you send your archive through an app, it is worth taking a look at the terms of use: what happens to the images, are they stored, may the provider reuse them? Especially if you later want to submit or sell photos on photo platforms, you should know which rights you have granted in advance.

Two more points: photos with recognizable people do not belong unchecked in third-party services, and neither do images from areas where photography was not allowed. And one more note for military spotters: most general models know military types much less well than airliners. If you want to identify transports or tankers, you are better advised to use specialist knowledge as in Identifying military transport aircraft.

Using AI as a learning aid

You get the most benefit from image recognition when you use it as a training partner. Instead of simply adopting the result, you ask yourself after every suggestion: what could I have seen myself? After a few weeks you recognize the common types at your home airport without an app – and you only need the AI for the difficult cases.

A good starting point is the guide Identifying aircraft types for beginners. If you are further along, you can take on the typical confusion pairs in a targeted way, for example with Telling Embraer E-Jets apart. And if you want to train your hearing, Identifying aircraft types by sound offers a completely different approach.

What to watch out for

• Family before variant. If the exact variant is unclear to you, “A320 family” or “737” is more honest than a guessed subtype.
• Take contradictions seriously. If the AI says something different from the registration, the registration wins – provided you read it correctly.
• Do not overestimate confidence values. If an app shows a high percentage, that only means the model is sure – not that it is right.
• Do not sink into the display at the spot. While you analyze a photo, the next aircraft rolls by. You can do the check calmly at home; Logging sightings quickly at the spot describes quick capture at the fence.
• Compare apps. Which spotting apps offer which features is shown in the plane spotting apps comparison.

Conclusion

AI image recognition is a useful tool in plane spotting, but not an oracle. It classifies aircraft roughly reliably and recognizes distinctive types, but stumbles exactly where it gets interesting for spotters: fuselage stretches, engine generations and rare variants. AR views in flight tracker apps are often more accurate in the sky because they evaluate position data instead of pixels, but have gaps of their own. The most reliable remains the combination of registration, fleet database and your own look at a few hard features.