A drone, being built, to photograph Indian ground and give the pictures away.
Before there was a bill of materials, there was a night, a laptop, and a very specific kind of frustration.
A data problem. Not a conspiracy. Fixable with an afternoon of engineering and about three months of fieldwork.
Most public computer-vision datasets used in India were collected somewhere else. That is not a failure of the researchers. It is a failure of what was available, and what was available came from a handful of well-funded labs in a handful of well-funded countries.
The result: models that score well on benchmarks and fail quietly in the field. A road classifier that has never seen a monsoon-flooded road. A land-cover model that reads a coconut grove as forest. A segmentation network trained on Dutch farms trying to parse a rice paddy outside Margao.
Nobody notices, because nobody is measuring. The benchmarks are foreign. The papers cite numbers from Berkeley and Zurich. Nobody asks what happens in Khordha or Kutch or, for that matter, on the road between Panjim and Canacona.
The fix is tedious, not complicated. Put a camera in the air over Indian ground. Document it properly. Publish the whole thing so the next model has somewhere to learn from.
Chosen for endurance and payload headroom - not for speed.
The 5-inch kit was the obvious starting point - cheaper, easier to carry, and well-proven in the freestyle world. I looked at it seriously. But a 5-inch is a freestyle airframe. It is built to move fast and crash well, not to hover steadily over a rice paddy for twenty minutes with a camera hanging off the front.
The 7-inch gives me inertia, lift, and calm. It will not win a race. It will hold a line in a crosswind and come home with the footage intact. For this project, that is what matters.
Nothing has been ordered yet. Prices below reflect previously discussed listings and will be replaced with confirmed supplier quotes before any purchase is made.
Nothing flies until it is trustworthy. Nothing publishes until it is clean.
Confirming the bill of materials, checking electrical compatibility, collecting real supplier quotations. Resolving the video-system question.
Parts arrive, bench build, soldering, wiring, firmware. Every step photographed and documented as it happens.
Hover tests, PID tuning, failsafe checks, endurance runs, payload verification. No capture flights until the aircraft is stable and predictable.
Permitted sites across varied terrain and lighting. Every session logged with conditions, altitude and camera settings.
Sorting, metadata, quality checks, privacy review, licensing. Then the repository and dataset go public.
Images. Metadata. Processing code. Build log. Example notebooks. Open licences.
A dataset is only useful if people can actually use it. That means no paywall, no API key, no application form. It means clean metadata, clear file naming, and a README that explains what was captured and why.
It also means being honest about limits. What the dataset supports depends entirely on what gets captured. I would rather under-claim now than over-claim later, when someone builds on top of it.
india-from-above/
├── README.md
├── LICENSE — MIT (code) · CC BY 4.0 (imagery)
├── docs/
│ ├── project-overview.md
│ ├── drone-build.md — bill of materials · wiring · tuning
│ ├── data-collection.md — sites · conditions · camera settings
│ └── privacy-and-safety.md
├── dataset/
│ ├── README.md
│ ├── metadata/ — location class · altitude · conditions
│ └── sample-images/
├── code/
│ ├── capture-processing/
│ └── dataset-tools/
└── research/
└── example-workflows/
Progress, setbacks, purchases, flight reports. Posted here first, then shared.
Target is ₹2,79,228 for the full build: drone platform, batteries, camera, and fieldwork across multiple sites. First contribution came in - ₹100.
Nothing ordered yet. Collecting real quotations for the 7-inch configuration so the ledger reflects actual prices. Working through electrical compatibility: Li-ion discharge rating, ESC headroom, connector types, camera mounting clearance on the LR7 frame.
Compared a 5-inch digital kit against a 7-inch long-range platform. 5-inch was cheaper and lighter, but it is a freestyle airframe. For slow capture with a GoPro on board, the 7-inch is the honest choice.
Going with the larger frame. Trade-off accepted.
Custom drone. Indian aerial imagery. Open publication. Privacy-conscious development. Target set at ₹2,79,228. Intended outputs: imagery, metadata, code, documentation.
A project that hides its open problems is selling something. These are ours.
The project started with a stated preference to avoid DJI. The current configuration still uses a DJI video system and goggles. Either the requirement softens with a clear justification, or those components are replaced. Decision required before purchase.
The component combination has not been independently verified or test-flown. Real flight time with a GoPro mounted is unknown until Phase 03. No endurance figures will be published until they are measured.
Applicable Indian drone rules, registration, and site permissions are being reviewed. Long-range hardware does not authorise beyond-visual-line-of-sight flight or flight in restricted areas. Nothing flies until this is settled.
The ₹28,990 allowance for fieldwork, storage, spares, shipping, and taxes is a planning figure - not a confirmed allocation. It will be replaced by a real cost sheet before purchase.
The target is ₹2,79,228 - hardware, batteries, camera, and the fieldwork needed to collect, organise and publish the dataset. Every output goes public.
Hi -
If you want to ask about the build, the aircraft, the dataset plan, the licensing, the fieldwork, or the AI research angle - this is the address. I read every email that lands here.
But I'm one person with a full-time build in front of me, so - proper questions only. Read the dossier first. If what you're asking is already answered in Chapter 2 or 4, I probably won't reply. Not because I don't care, but because I'd rather spend that hour on the drone.