🌡 @smartplant/thermal

Reads the canopy with a thermal camera: where it is transpiring, and where it has stopped.

A plugin for smartplant — a bridge between AI and plants.

Install

npm install smartplant @smartplant/thermal

Use

import { createPlant } from 'smartplant'
import thermal from '@smartplant/thermal'

const plant = await createPlant( { name : 'Ivy', species : 'Ficus', sensor : 'mock', ai : { provider : 'ollama' } } )
await plant.use( thermal )

await plant.attachSensor( {
  driver : 'thermal',
  frame  : async () => readLepton(), // { width, height, data } — °C per pixel
} )

console.log( await plant.plugin( 'thermal' ).stress() )
console.log( await plant.plugin( 'thermal' ).evenness() )

Why a camera rather than a leaf clip

A leaf cools itself by evaporating, so a working leaf sits below air temperature and one that has shut its stomata drifts up toward it. That gap is water stress, hours before anything looks wilted.

A contact probe reports whichever leaf it was clipped to. A frame reports the whole canopy — and half a plant in trouble looks identical to all of it through a clip. A blocked vessel, a root problem on one side or a branch sitting in a draught all show as unevenness long before they show as anything visible.

An uncalibrated camera does not know the temperature

A Lepton-class sensor is accurate to roughly ±5 °C absolutely and to a few hundredths within one frame. Emissivity, the reflected background, distance and the sensor's own temperature all move the absolute reading, and a leaf is not a blackbody.

So nothing here is published as leafTemperature, and every number is a difference. It also means the air temperature has to come from an ordinary thermometer: reading it off the background of the same frame would cancel the sensor's error out of both numbers and make their difference look perfect while meaning nothing. Without one, stress() refuses.

API

stress()

{ known, index, gap, why }. Zero is a plant cooling itself freely; one is a plant that has closed up. Indoors the whole scale is compressed into about three degrees, so half a degree moves it a long way.

evenness()

{ even, spread, why }. The finding worth owning a camera for.

map()

{ rows, min, max, relative }. False colour for a panel, scaled between the coolest and warmest pixel in the frame — relative on purpose.

Hardware

No hardware library is imported. frame is a callback returning { width, height, data }; whatever produced it is yours to choose.

One case worth knowing: a plant that has stopped transpiring altogether sits at room temperature and disappears into the wall behind it. The segmentation reports that as ambiguous rather than as a plant at room temperature — a plant vanishing from a thermal image is a finding, not a failure to find one.

Test

npm test

Runs against the mock sensor and mock AI provider — no hardware, no API key, no network.

📜 License

MIT — see LICENSE.