Scene 01 — Case · Samsung Innovation Campus · 2024—25

PlantSense

IoT Acoustic Plant Stress Detection

Hearing a plant run out of water, on under €10 of hardware.

A plant short of water pulls harder on the water in its conducting tissue. Now and then one of those columns breaks under the tension, and the break makes a sound — far above anything a person can hear. Nobody had to discover that. What was still missing was an instrument cheap enough to leave next to the plant, listening.

FFT
Custom analogue circuit
RP2040
Embedded

Kind: Applied project, Samsung Innovation Campus
Period: 2024—25
Sensing: Ultrasonic acoustic emission from cavitation
Hardware: Raspberry Pi Pico W (RP2040) and a custom piezo circuit
Cost: Under €10 of hardware
Result: 173.5 vs. 2.5 events/hour

Channel A — water-stressed

347events

173.5 per hour

Channel B — healthy control

5events

2.5 per hour

69×separation margin
173.5 ÷ 2.5 = 69.4

Figure 1 — the acoustic emission record. Two channels, one time axis, one amplitude scale: the difference between them is event density, not drawing. The window is two hours because the measured rates are hourly, and two hours makes both counts whole — 173.5 × 2 = 347 and 2.5 × 2 = 5. Arrival times are the order statistics of a fixed-seed uniform draw, which is the correct arrival distribution for a Poisson process with an observed count; event amplitudes are illustrative. Schematic, generated deterministically at build time; not a capture.

Scene 02 — Problem

The measurement was solved. The instrument was not.

Detecting plant water stress from ultrasonic cavitation is established science. It is not a hypothesis I was testing and not a result I improved. The physics has been understood for a long time, and if you put the right equipment next to a thirsty plant, it works.

The catch is in "the right equipment". The technique lives in laboratories, on instrumentation costing thousands of euros, in rooms arranged so that the instrument hears the plant and very little else. That is a perfectly good place for a technique to live if the question is whether the effect is real. It is a useless place for it to live if the question is whether a greenhouse should turn on a valve.

So the open problem was never can a plant be heard. It was whether a plant can be heard by something you could leave in the pot and forget about — and that is a cost problem, not a physics problem. Nothing in this project asks the science to be different. The entire redesign is on the other side.

Two columns

Already settled: That a water-stressed plant emits ultrasonic acoustic events, and that counting them tells you something about its water status.
Still open: That anyone outside a laboratory could afford the thing doing the counting.

Rebuilt for under €10.

The laboratory side of that comparison is a class of equipment, not a quotation I obtained. The only price here I can stand behind is the one I paid.

Scene 03 — Approach

Build the instrument down

A piezo element, a custom analogue circuit around it, a Raspberry Pi Pico W, and an FFT. That is the whole instrument, and every one of those four pieces is doing the job of something that normally sits on a bench.

The piezo element is the part that cannot be cheapened away — something has to turn a mechanical click into a voltage. Everything after it can be, and the conditioning stage is where that happens: a discrete circuit designed around this particular element so that what reaches the microcontroller is legible rather than merely present.

The FFT is the piece that buys the most. A laboratory does not need one, because it can afford quiet: it isolates the plant physically and then reads whatever the transducer returns. A sensor that has to live in a greenhouse cannot buy quiet, so it buys separation instead — the plant's events sit in a band the room does not occupy, and the transform is what makes that separation usable.

What leaves the device is not a recording. It is a count of events per hour, which is the number the science actually reads and the only number anyone downstream needs.

The event

mechanism

Schematic of the mechanism, not to scale and not a capture. The device does not see any of this. It counts the sound the break makes.

Figure 2 — separation by band, schematically. The horizontal axis carries no numbers on purpose: the band edges for this device were not measured, so putting frequencies on the axis would dress an argument up as a result. The one marked position is the ceiling of human hearing, which is textbook, not a property of this instrument. What the figure does claim is the shape of the problem and the shape of the answer: ambient noise owns the low end, the plant's events do not live there, and an FFT is what turns that fact into a usable separation. Illustrative, not measured output.

The signal chain

01Piezo elementA mechanical click becomes a voltage. This is the part that costs almost nothing and does the part nobody can do differently.
02Conditioning stageA custom analogue circuit built around that element, so what comes out of it is something a converter can actually read.
03Raspberry Pi Pico WRP2040. The whole acquisition side of a laboratory bench, in the pot with the plant.
04FFTThe plant band is separated from ambient noise here — which is why the room around the sensor does not have to be quiet.
05Event countThe output is not a waveform to be read later. It is a number of events per hour, leaving over the network.

Scene 04 — Obstacle

Everything the bench gave you, you now owe

Taking the price out of an instrument is not a discount. Every line item you delete was doing something, and the work does not disappear with the cost — it moves somewhere else in the design, usually somewhere less comfortable.

Four things had to be replaced rather than removed. The left column is the class of equipment the technique normally assumes; it is a description, not a quotation. The right column is what stands in its place here.

None of these substitutions changes the quantity being measured. They change where the measurement is allowed to stand — which was the point.

Transducer
A calibrated ultrasonic transducer, specified and characterised.
A piezo element, with a conditioning stage designed around it to make its output legible.
Acquisition
A bench digitiser, and a workstation behind it to hold the capture.
A Raspberry Pi Pico W wired to that circuit, doing both jobs.
Isolation
A room arranged so that the instrument hears the plant and little else.
An FFT — so the room does not have to be arranged at all, only the band has to be chosen.
Reading
A recording that somebody opens afterwards and interprets.
An event count per hour, sent over the network from wherever the plant is standing.

Scene 05 — Result

173.5 against 2.5

Under continuous acoustic monitoring, the water-stressed plant produced 173.5 events per hour. The healthy control produced 2.5. Stressed and healthy separated with a decisive margin, and the margin is the reason a cheap instrument is enough.

173.5/h

Events per hour from the water-stressed plant.

2.5/h

Events per hour from the healthy control.

69×

Separation margin — arithmetic, 173.5 ÷ 2.5 = 69.4.

< 10€

What the hardware for one unit cost.

Rates, on one scale

control ÷ stressed

Water-stressed plant173.5 /h

Healthy control2.5 /h

Both bars are drawn against the stressed rate, which is a real denominator: the control sits at 1.4% of it — arithmetic, 2.5 ÷ 173.5. The second bar is almost invisible, and that is not a rendering problem. It is the result.

Why the margin matters more than the precision

A detector needs precision it can afford. When the two classes it has to tell apart are nearly two orders of magnitude apart in event rate, the instrument does not have to be delicate — it has to count, and it has to not count the room. That is exactly the trade this build makes, and it is why the cost could come down without the decision getting worse.

What is not here: a false-positive rate, a per-event classification accuracy, a trial across species, seasons or field sites. Those were not measured, and the two rates above do not stand in for them.

98.6%

Share of all events counted across both channels that came from the stressed plant — arithmetic, 173.5 ÷ 176.0.

1.4%

The healthy control's rate as a share of the stressed plant's — arithmetic, 2.5 ÷ 173.5.

PlantSense measurements and the conditions they were measured under
Metric	Value	Measured under
Stressed plant	173.5 events/h	Water-stressed plant under continuous acoustic monitoring
Healthy control	2.5 events/h	Healthy control plant, same instrument, same conditions
Separation margin	69 ×	Arithmetic — 173.5 ÷ 2.5 = 69.4, reported as 69×
Hardware	< 10 €	What the hardware for one unit cost
Controller	RP2040	Raspberry Pi Pico W, with a custom piezo circuit and conditioning stage
Separation	FFT	Plant emission isolated from ambient noise; band edges not recorded
Programme	2024—25	Samsung Innovation Campus

Scene 06 — Close

What this case is evidence of

I work on measurement, and on what depends on it.Ali Batuhan Özsürmeli

The technical content here is a piezo element, an analogue front end and an FFT. The transferable part is the choice of what to redesign. The measurement was already solved and I did not improve it; what was unsolved was the instrument, and an instrument problem is a cost problem. So the useful question was never "can this be detected" but "what is the cheapest thing that can detect it, and what does that thing now owe".

A laboratory technique that reaches a greenhouse is worth more than the same technique measured more precisely on a bench. That is the same judgement that runs through the rest of this work: a result is only as good as the conditions it survives.

Techniques

applied

FFT
Custom analogue circuit
RP2040
Embedded

Programme: Samsung Innovation Campus, 2024—25
Metric: 173.5 vs. 2.5 events/hour