Work in progress

Pitchfork

A handheld camera for measuring plants.

Colour, red, near-infrared and thermal in one shot.

Pitchfork photographs a plant with four cameras at once and measures the leaves alone, leaving soil and plot labels out of the numbers. It gives you NDVI, a standard index of how healthy the leaf tissue is, and leaf temperature. We are building it for plant researchers and breeders who measure single plants and small plots. Today it is a working prototype on our bench.

How it works ↓ Where it stands ↓
Colour, as the eye sees it. Used to frame the shot and check focus. Illustration, not a capture.

What it measures

Temperature and NDVI, from leaves only.

A plant short of water closes the pores in its leaves, stops cooling itself and warms up before it wilts or changes colour. A thermal camera can see that. It also sees warm soil, stakes and plot labels, and in a sparse canopy those pull the average up. So Pitchfork first works out which pixels are leaf, then measures temperature and NDVI on those pixels only. It works close up, at 0.3 to 1 m, where single leaves are resolved.

Colour

For framing the shot and checking focus, and as a record of what the plant looked like on the day.

Red

Chlorophyll absorbs red light. A black-and-white camera behind a red filter measures how much the leaf sends back.

Near-infrared

Healthy leaf tissue reflects near-infrared strongly. A second camera of the same kind sits behind a near-infrared filter. Together with red it gives NDVI for every pixel.

Thermal

A FLIR Lepton 3.5 core, 160 × 120 pixels. It is radiometric: each pixel is a temperature in degrees.

How a capture works

From a button press to numbers.

  1. 1
    Pitchfork aimed at a plant

    Aim

    The 5-inch touchscreen shows a live picture and switches between colour, thermal, red, near-infrared and live NDVI. Hold Pitchfork 0.3 to 1 m from the plant and press the capture button.

  2. 2
    Four images of the same plant: colour, red, near-infrared and thermal

    Four pictures

    One press takes a picture from each camera. The thermal and colour cameras run on their own clocks, so their nearest frames are used. Red and near-infrared are lined up with each other by a calibration fitted on a test pattern.

  3. 3
    The plant before and after masking: soil, stake and label removed, leaves kept

    Leaves only

    Pitchfork sorts leaf pixels from soil, stakes and labels. For a new crop you can train a new leaf model in Hayloft, the companion app, by tracing leaves on your own captures.

  4. 4
    An example readout of leaf-only NDVI, leaf temperature and leaf cover

    Numbers

    NDVI, leaf temperature and leaf cover from leaf pixels only, plus a stress map that combines NDVI with how much warmer the leaves are than the air. The raw bands are saved as NumPy arrays with JSON metadata, so every number can be recomputed in Python or R.

Filters and indices

Swap the filters, change the index.

An index is a simple formula on two pictures taken in different bands of light. NDVI is the best known: near-infrared minus red, divided by near-infrared plus red. Healthy leaves reflect a lot of near-infrared and little red, so they score high and soil scores low. Pitchfork's two band cameras sit behind filters you can swap. A red-edge filter in place of the red one gives NDRE, and a green one gives GNDVI. In principle any index built from two bands works.

NDVI to start with.

The planned pair is red at 650 to 660 nm and near-infrared at 850 nm, each 40 to 50 nm wide. We picked wide bands because a filter's band shifts towards blue at the edges of a wide-angle picture, and a narrow one would miss its target there.

Your bands.

Which camera carries which filter, and at what wavelength, is a setting on the device. A leaf model remembers the bands it was trained on, so new filters need a new leaf model from Hayloft.

Two at once, in principle.

Pair each filtered camera with the red, green or blue channel of the colour camera and one shot gives two indices. That needs the colour camera lined up with the band cameras, which we have not done yet.

What's inside

Off-the-shelf parts on our own mainboard.

The two band cameras, the thermal core and the colour camera share one sensor plate at the front. Behind them sit a single-board computer that runs the screen and does the maths, a smaller one that runs the colour camera, and a microcontroller that reads the sensors and fires the band cameras. The sensors measure distance, tilt, air temperature and humidity, and light on the plant and from the sky. All of it plugs into a mainboard we designed ourselves, which we are now testing one function at a time.

Everything, battery included, goes into one tube: light-coloured outside, black inside, with the screen tilted at the back and a rail underneath for a handle or a tripod. That tube is still being drawn.

The companion app

Hayloft runs in your browser, offline.

Captures come off Pitchfork on a USB stick and open in Hayloft, our companion app (named after where the harvest goes once the pitchfork is done with it). Every file is decoded and every map drawn on your own computer. Nothing is uploaded and there is no account. Hayloft is a single HTML file that runs straight from the stick. It also installs as an app and runs on Pitchfork's own screen. Leaf models and settings go back to the device on a stick, and the device checks them and shows what would change before you apply them.

Pixel illustration: Hayloft in a browser window drawing a leaf stress map, with a crossed-out cloud for no upload.

No cloud anywhere.

Your trial data stays on the machines you use. Pitchfork carries its own OpenStreetMap map of north-west Europe, so the map view works in a field with no signal.

Made for trials.

Load a trial plan and the device walks you from plot to plot, so captures arrive labelled. Hayloft compares treatments with the plot as the replicate (Welch's t-test and ANOVA, checked against SciPy) and prints a report.

Same maths on both sides.

Hayloft recomputes captures with the device's own processing. Its tests run on a real capture from the device: the leaf mask comes out identical pixel for pixel, and the NDVI statistics agree to three decimals.

Where it stands

What works, and what doesn't yet.

Pitchfork is a working prototype. The electronics run open on our bench, wired to the new mainboard, and the cameras sit in an early test housing. It has not been out on a real crop yet, and that is the biggest test still ahead.

  1. Four cameras running
  2. Bands aligned
  3. Live NDVI
  4. Hayloft app
  5. Hardware trigger
  6. Mainboard
  7. Final filters
  8. Enclosure
  9. Battery
  10. Reference paddle
  11. Field test
  12. Early units

Works on the bench

  • All four cameras run, and the touchscreen switches between them and a live NDVI picture at about four frames a second.
  • Red and near-infrared line up to about 1.4 pixels after calibration on a test pattern.
  • One electrical pulse fires both band cameras: 1,000 pulses gave 1,000 frames on each, never more than 0.023 ms apart.
  • On the new mainboard so far: power, the link to the computer, tilt, air temperature and humidity, distance, both light sensors, the capture button and the camera trigger.
  • Hayloft, with trial statistics, the report and the offline map.

Not yet

  • Captures still use a software trigger, where the two band cameras can be up to 27 ms apart. A leaf moving in the breeze can land on different pixels in the two pictures.
  • Neither NDVI nor temperature is checked against a reference yet. NDVI is computed from raw camera counts, and the thermal core on its own is only accurate to about ±5 °C.
  • The thermal camera's video link drops out now and then on the bench. It restarts itself within seconds, and we have not found the cause.
  • No GPS yet. The module is chosen but not wired.

Next

  • Finish testing the mainboard.
  • Move capture onto the hardware trigger.
  • Fit the final filter pair. The bench cameras run with spare filters, or none.
  • Draw and print the enclosure, and fit the battery.
  • Build the reference paddle: a white patch for reflectance and a temperature-controlled plate for the thermal camera.
  • Take it out to a real crop.

Would you like to try an early unit once there is one? Email and tell us what you grow.

Specification

What we are aiming for.

Targets for the first units, from the current design. Some will change as we test.

Bands
Colour, red, near-infrared and radiometric thermal
Band cameras
Two global-shutter monochrome cameras, 1456 × 1088 pixels, behind swappable band-pass filters
Filters
Red at 650 to 660 nm and near-infrared at 850 nm, each 40 to 50 nm wide
Indices
NDVI as standard; in principle any two-band index, and two at once with colour channels
Thermal
FLIR Lepton 3.5, radiometric, 160 × 120 pixels
Working distance
0.3 to 1 m
Form
Handheld, with a rail underneath for a handle or a tripod
Screen
5-inch touchscreen
Battery
About 3 to 4 hours
Position
GPS with every capture, good to a few metres in the open
Data
Raw bands as NumPy arrays plus JSON metadata, exported to a USB stick
Software
Hayloft, in the browser, offline, nothing uploaded

Questions

Short answers.

Can I buy one?

Not yet. Pitchfork is still a prototype, so there is no price and no date.

How accurate is the thermal reading?

On its own, not accurate enough. The Lepton 3.5 core is typically accurate to about ±5 °C, and a plant short of water warms up by 1 to 3 °C. That is why we are building the reference paddle, with a temperature-controlled plate the camera measures at start-up. Until that is built and tested, treat absolute temperatures with caution.

Does it replace a drone?

No. A drone maps whole fields from above, while Pitchfork looks at single plants from 0.3 to 1 m, close enough to tell leaf from soil.