Find drone footage fast: from DJI_0001.MP4 to a searchable library

Aerial footage has a specific and infuriating property: it is the most valuable material in your library and the hardest to find. A hundred drone clips in a thumbnail grid look like a hundred rectangles of green, blue and grey. The one you are thinking of, the pull-back over the coast road at the moment the light broke, is in there somewhere, and thumbnails will not help you.

Here is why the usual approaches fail on drone footage specifically, and what to do instead.

Why thumbnails fail for aerial

For a ground shoot, a contact sheet works. You see faces, rooms, framing, and you recognize the moment. For aerial material, three things break at once:

Low visual variance. From 80 metres up, most coastlines look like most coastlines. The differences that matter to you, the exact angle, the light, the moment the boat entered the frame, are not visible in a 200 pixel thumbnail.

The first frame is the worst frame. Drone shots usually start static and become interesting as the move develops. Your thumbnail is almost always the least representative part of the clip.

Volume. Drone pilots shoot long. Ten minutes of battery, four takes of the same reveal, three altitudes. You end up with many near-identical files that differ only in the part you cannot see.

The conclusion: for aerial footage, text is a better index than pictures. Which means the work is getting text attached to each clip.

What DJI actually gives you

Before adding anything, use what the drone already wrote.

Filenames. Older models write DJI_0001.MP4. Newer ones write a timestamped form:

DJI_20260613_153045_0001_D.MP4

That is date, time and sequence, which is genuinely useful: it means the file itself knows when it was shot, so grouping flights by session works without any extra tooling.

SRT sidecars. This is the part almost everyone deletes. DJI writes a .SRT file alongside many clips containing per-second telemetry: GPS latitude and longitude, altitude, ISO, shutter, focal length, timestamp. It uses the subtitle format as a container, which is why it looks like a subtitle file and gets thrown away.

Open one in a text editor and you will see entries like coordinates and altitude for every second of flight. That is the only record you will ever have of where a shot was taken. Two things to do with it:

• Keep it with the clip, forever. It is a few kilobytes.
• Pull the first coordinate into your naming or notes. A location name in the filename beats a coordinate you have to look up, but a coordinate beats nothing.

Panorama and hyperlapse output. These land as separate stills or as a series. Give them their own folder immediately, or they will pad out every future clip search with hundreds of JPEGs.

The three axes that make aerial searchable

Ground footage needs subject and scene. Aerial footage needs a third: the move. Tagging along all three is what turns a library into something you can query.

Subject and location type. Coastline, forest, city, farmland, industrial, harbour, mountain, road, river, single building. Keep this list short and reuse it across projects, or it stops being comparable.

Time and light. Golden hour, blue hour, midday, overcast, night. For aerial this is not a nice-to-have. It is the first filter you apply when matching a shot to a cut.

The move. This is the one people forget and then search for constantly:

• orbit
• reveal, usually a rise or a push past an obstruction
• top down, sometimes called nadir
• fly through
• pull back
• descent
• tracking, following a subject

“I need a pull-back over water at golden hour” is a real request you will make of your own archive. Without the move recorded, it is unanswerable except by scrubbing.

Where to write it down

Three places, in order of durability.

The filename. Visible everywhere, survives everything:

DJI_20260613_0042_Amalfi-Coast-Road-Pullback_Golden-Hour_16x9_4K_30FPS.mp4

Finder tags. On macOS these are indexed by Spotlight, which means a search from any application’s open dialog finds them. Tag with the move and the light: orbit, golden-hour, coast. A short controlled list, applied to everything, beats a rich list applied to some things. More on that in using Finder as a footage browser.

A sidecar text file. For the things too long for a filename: the location name, the flight notes, the reason a take was rejected. Spotlight indexes text files, so this content becomes searchable too.

What not to do: put it only in your NLE’s project. Premiere and Resolve metadata is excellent and completely invisible from outside the project file. When you come back in a year, you will be searching the drive, not the project. That is the case laid out in finding clips without opening Premiere.

Generating the descriptions instead of typing them

Describing 200 aerial clips by hand is an evening. A vision model does the same pass unattended, and aerial footage is a good case for it, because what you want to search for is exactly what is visible from above.

What works well:

• Location type: coast, forest, city, farmland, desert, snow, industrial
• Water versus land, built versus natural
• Time of day and light conditions
• Camera move, in most cases
• Obvious subjects: a boat, a car on a road, a stadium, a bridge

What does not work:

• Naming the specific place. A model may guess “Mediterranean coastline”, which is useful, or guess a specific city wrongly, which is worse than nothing. Pair it with the GPS from the SRT file.
• Altitude. It is in the telemetry, not in the picture.
• Airspace, permissions, flight logs. Not visible, ever.

Realistic expectation: generated descriptions get you 80 percent of the way, and the last 20 percent, the place name and the project context, is a quick pass by the person who was there. That is still an evening saved per shoot.

For material under NDA or with recognizable private property, an on-device model does this without frames leaving your machine, at slightly coarser quality.

A ten second retrieval, from the top

The test of any system is retrieval. Here is what it looks like when the pieces are in place, and you need a golden hour pull-back over a coast road:

1. Spotlight or a Finder search window.
2. tag:golden-hour tag:coast pullback
3. Six results, all with filenames that describe them.
4. Space bar to preview.

No NLE, no catalog app, no scrubbing. That is the whole point of the exercise, and it works because every step happened at ingest, once, instead of at retrieval, repeatedly.

Automating the pass

Cliptag does the ingest pass on macOS: it analyzes each clip’s frames and audio, writes a descriptive filename, applies Finder tags, and files everything by media type and date. Drone material, stills from the same flight and any audio you recorded on the ground can go in as one drop.

Worth knowing for drone work specifically:

• Standard H.264 and H.265 from DJI, including 10-bit, are read directly.
• The .SRT telemetry files are not video and are left untouched next to the clip, which is exactly what you want.
• Original files are never deleted or modified, and every rename and move is logged so it can be reconstructed.

The on-device mode on Apple Silicon is free and unlimited, and the free plan includes 25 cloud analyses a month for the material where the extra precision is worth it.