A normal camera collects three numbers per pixel — red, green, blue — roughly what a human eye reports. A hyperspectral camera collects a hundred or more, each a narrow slice of the spectrum a few nanometers wide, so every pixel carries a reflectance curve rather than a color. That curve is a material signature. Chlorophyll absorbs strongly near 680 nm and reflects hard just past 700 nm; water absorbs near 970 and 1,450 nm; minerals, plastics, and coatings each have their own bands.
So hyperspectral imaging answers questions color imaging cannot: which of these visually identical plants is nitrogen-stressed, which roof section holds moisture inside the membrane, which tailings pile contains the mineral of interest. Flown low from a drone, it answers at a spatial resolution satellites cannot approach.
Multispectral vs Hyperspectral: The Distinction That Matters Commercially
The two get conflated constantly, and the difference decides the budget.
| Multispectral | Hyperspectral | |
|---|---|---|
| Bands | 4–12 discrete bands | 100–400+ contiguous narrow bands |
| Spectral width | 20–50 nm each | 2–10 nm each |
| Payload mass | 150–700 g | 0.6–4 kg with optics and storage |
| Payload cost | $4,000–$20,000 | $25,000–$120,000 |
| Data per flight | Gigabytes | Tens to hundreds of gigabytes |
| Best for | Known indices like NDVI and NDRE | Unknown or subtle signatures, material identification |
If your application is served by a published index — vegetation vigor, canopy cover, simple water stress — buy multispectral and stop. It costs a fraction as much, the data is manageable, and the processing pipeline is mature. That is the right answer for most of the work described in agricultural drones for precision farming. Hyperspectral earns its cost when you do not yet know which wavelengths matter, when you need to discriminate between things that share a broad band, or when you are building a classifier from scratch.
How the Sensors Work, and Why It Constrains the Flight
Push-broom (line scan)
The dominant architecture. The sensor images a single line across the ground and disperses it through a spectrograph onto a 2D detector — one axis spatial, the other wavelength. The second spatial dimension comes from the aircraft's own motion, which makes the datacube a function of the flight path: every attitude wobble smears the image. You need a stabilized mount, a high-rate IMU logging attitude at the frame rate, and constant-speed flight lines, or you get stretched and compressed swaths that are painful to rectify.
Snapshot
Captures a full 2D image with spectral information in one exposure, using a filter mosaic or tiled interferometer. Far more tolerant of aircraft motion and much easier to fly, at the cost of fewer bands, lower spatial resolution, or both. For stop-and-stare inspection rather than survey transects, snapshot sensors are often more practical.
Wavelength range
VNIR sensors cover roughly 400–1,000 nm on silicon detectors and are the affordable option — enough for vegetation, pigments, and many surface features. SWIR sensors cover roughly 1,000–2,500 nm, need a cooled InGaAs detector, and cost three to five times as much, but that is where moisture, hydrocarbons, plastics, and most minerals show their diagnostic features. Range is the first and most expensive decision. On detector selection generally, see choosing an image sensor.
Calibration Is Most of the Work
Raw hyperspectral output is digital counts, meaningless across flights, times of day, and cloud conditions. Turning counts into reflectance is what makes results comparable, and skipping it is why so many pilot projects produce pretty images and no decisions.
- Radiometric calibration from the sensor vendor, plus dark-current frames captured each flight.
- Reflectance panels laid in the scene — calibrated Lambertian targets at known reflectance levels, imaged before and after the flight.
- A downwelling irradiance sensor on top of the aircraft, logging incoming light continuously so that a passing cloud does not become a false anomaly.
- Atmospheric correction, lighter than for satellite data but not negligible over long transects.
- Geometric correction and georeferencing, which for a push-broom sensor requires tight fusion of the attitude and position solution with each scan line. RTK or PPK GNSS is effectively mandatory — this is the same positioning discipline described in drone photogrammetry and survey-grade accuracy.
Fly near solar noon, keep sun angle consistent between flight lines, and avoid partly cloudy conditions entirely. Hyperspectral survey is far more weather-constrained than ordinary aerial imaging, and that belongs in the operational plan and the customer's expectations.
Integrating It on the Aircraft
The payload is heavy, hungry, and produces data faster than most onboard links can move it. Practical integration issues:
- Mass and balance. A 2 kg payload on a survey multirotor means a larger airframe or a fixed-wing VTOL. Endurance drops, and endurance sets how many acres per flight you can deliver, which sets your unit economics.
- Power. A SWIR sensor with a cooler can draw 20 to 60 W continuously — plan the pack around it rather than treating it as a rounding error.
- Storage and throughput. Hundreds of megabytes per second in some configurations. Onboard NVMe storage and a physical card swap after landing, not a wireless download.
- Timing. Every frame must be timestamped against the GNSS and IMU solution to sub-millisecond accuracy, or georeferencing degrades.
- Vibration and thermal drift. A spectrograph is an optical alignment; vibration blurs the spectral axis and temperature shifts the calibration. Isolation and thermal stability matter more than for a normal camera — the same discipline as in gimbal and vibration isolation design.
The Data Problem Is the Real Project
A hyperspectral flight produces a datacube, not a picture. Somebody has to reduce it into a decision: a stress map, a mineral classification, a pass/fail. That means dimensionality reduction, a library of reference spectra or a trained classifier, and ground truth collected the same day — soil samples, tissue tests, core assays. Budget as much engineering effort for the processing chain and validation as for the aircraft. Customers do not buy datacubes.
Where It Pays Today
Working applications include crop disease and nutrient status ahead of visible symptoms, weed species discrimination for targeted spraying, mineral mapping and tailings characterization on sites like those in mining and quarry surveying, water quality monitoring, pipeline and tank leak detection in the SWIR, roof moisture surveys, and waste stream sorting. In several of these a thermal payload answers the same question far more cheaply, so always check whether a thermal camera will do before committing to a spectrograph.
Whichever sensor you land on, the integration, calibration workflow, and processing pipeline are the deliverable — and they dominate the budget, as our breakdown of drone development cost reflects. Projects House builds sensor-carrying unmanned systems for US clients from payload selection through fielded data products. If ordinary imaging cannot solve your detection problem, tell us about it through our contact form.