Audio

How many clicks

In the 15 minutes following a stimulus, mechanically stressed plants produced 105 clicks/hour counted automatically, and 41 clicks/hour counting only the ones confirmed by eye. Watered plants reached 324 and 134 clicks/hour. Unstressed plants stayed at 9.2 and 2.5, and empty rooms at 2.5 and 0.

Every rate is measured over the same 15 minutes of recording, so only recordings that run at least that long are included — 14 mechanically stressed, 7 watered, 31 unstressed and 8 empty-room. A recording that stops after a minute would turn three clicks into 170 per hour, which says more about its length than about the plant.

Compared against the unstressed plants' own first 15 minutes — the strictest control we have, since those recordings are handled at the start exactly like the others — mechanical stress raises the rate 11.5 times (95% confidence interval 8.9–15) and watering 35 times (28–46). Both windows are the first 15 minutes, on each side of the comparison.

Restricted to the recordings a person reviewed, and counting only the clicks they confirmed, the same window gives 16× for mechanical stress (95% confidence interval 6.8–52) and 54× for watering (22–170). Those intervals are wide because this comparison rests on far less material — 9 and 2 stressed recordings against 8 controls carrying 5 confirmed clicks in total — so we quote the automatic figures as the headline and these as the check that agrees with them.

By species under mechanical stress: Ferocactus 125 clicks/hour (6 recordings), Aloe vera 66 (6), Kalanchoe 292 (1) and Spathiphyllum 32 (1). After watering: tomato 582 (2 recordings), strawberry 278 (2), Aloe vera 252 (1), Kalanchoe 240 (1), Spathiphyllum 56 (1). The number of recordings behind each figure is shown beside it, because several species rest on a single session.

The clearest control result: across 7.3 hours of empty-room recording reviewed click by click, not a single click was confirmed.

When they happen

The response is immediate and short-lived. Of every click confirmed by eye in a mechanically stressed recording, 93% arrive within the first two minutes — and the stimulus itself is applied 30 to 60 seconds in. Counted automatically, which also picks up events a reviewer would reject, the share is 68% for mechanical stress and 78% for watering.

Unstressed plants show nothing of the kind. Only 27% of their clicks fall in the first two minutes, and 8 of the 14 unstressed recordings produce no click at all in that window — against 1 of 21 stressed recordings. Their emissions are spread across the whole session: half an hour in, the median unstressed recording has still delivered only 58% of its clicks, while a stressed one passed 70% before the two-minute mark.

Afterwards the plant falls quiet again: past the first 15 minutes, stressed recordings drop to about 0.4 confirmed clicks/hour — indistinguishable from an unstressed plant.

Detecting stress automatically

Counting clicks afterwards is one thing; raising an alarm while the plant is still emitting is another. Our detector keeps a sliding three-minute window over the incoming clicks and signals a stress event when the count crosses a threshold. That threshold is calibrated only on control recordings, chosen to tolerate at most 0.1 false alarms per hour, and never on the stressed recordings it is then tested against.

Tested one recording at a time, leaving that recording out of the calibration, it detects 20 of 22 stressed recordings and raises 4 false alarms in 44.7 hours of control recording (0.09 per hour; 95% upper bound 0.21). The separation between stressed and control recordings corresponds to an area under the ROC curve of 0.98, and the calibrated threshold came out identical in every single fold. The two misses are quiet Aloe recordings with 5 and 7 clicks in total, which no threshold would catch.

What we do not claim

These results are indoor only. Outdoors, environmental sounds recorded with no plant present produce more detections than recordings of plants do, so our algorithm is not yet reliable in that setting and we report no outdoor rates.

A burst of environmental false positives is, to a burst detector, indistinguishable from a burst of plant clicks. This is a property of the method, not a bug to be tuned away, and it is why a person reviewed a third of the recordings by hand.

While analysing the corpus we found a narrowband 66 kHz interference tone contaminating nine recordings — a fixed, low-amplitude line, unlike the broadband clicks of a plant. Those recordings are identified and excluded before any number above is computed. Finally, the classifier was trained on part of this material, so the automatic counts on those recordings are optimistic; the counts confirmed by eye are not.