Article

Sounds emitted by plants under stress are airborne

and informative
Graphical abstract Authors
Itzhak Khait, Ohad Lewin-Epstein,
Recording ultrasonic plant sounds
Raz Sharon, ..., Nir Sade, Yossi Yovel,
Lilach Hadany

Correspondence
[email protected]

In brief
Plants emit species- and stress-specific
airborne sounds that can be detected in
acoustic chambers and greenhouses.

plant sounds

Highlights
d Plants emit ultrasonic airborne sounds when stressed

d The emitted sounds reveal plant type and condition

d Plant sounds can be detected and interpreted in a

greenhouse setting

Khait et al., 2023, Cell 186, 1328–1336

March 30, 2023 ª 2023 The Authors. Published by Elsevier Inc.
https://doi.org/10.1016/j.cell.2023.03.009 ll
 ll
OPEN ACCESS

Article
Sounds emitted by plants under stress
are airborne and informative
Itzhak Khait,1,6,9 Ohad Lewin-Epstein,1,7,8,9 Raz Sharon,1,2 Kfir Saban,1 Revital Goldstein,1 Yehuda Anikster,1
Yarden Zeron,1 Chen Agassy,1 Shaked Nizan,1 Gayl Sharabi,1 Ran Perelman,1 Arjan Boonman,3 Nir Sade,1,4
Yossi Yovel,3,5,10 and Lilach Hadany1,5,10,11,*
1School of Plant Sciences and Food Security, The George S. Wise Faculty of Life Sciences, Tel-Aviv University, Tel-Aviv, Israel
2School of Mathematical Sciences, Tel-Aviv University, Tel-Aviv, Israel
3School of Zoology, The George S. Wise Faculty of Life Sciences, Tel-Aviv University, Tel-Aviv, Israel
4The Institute of Cereal Crop Improvement, The George S. Wise Faculty of Life Sciences, Tel Aviv University, Tel Aviv, Israel
5Sagol School of Neuroscience, Tel Aviv University, Tel-Aviv, Israel
6Present address: Centure Applications (Greeneye Technology), Tel Aviv, Israel
7Present address: Broad Institute of MIT and Harvard, Cambridge, MA, USA
8Present address: Center for Computational and Integrative Biology, Massachusetts General Hospital and Harvard Medical School, Boston,

MA, USA
9These authors contributed equally
10These authors contributed equally
11Lead contact

*Correspondence: [email protected]
https://doi.org/10.1016/j.cell.2023.03.009

SUMMARY

Stressed plants show altered phenotypes, including changes in color, smell, and shape. Yet, airborne sounds
emitted by stressed plants have not been investigated before. Here we show that stressed plants emit
airborne sounds that can be recorded from a distance and classified. We recorded ultrasonic sounds emitted
by tomato and tobacco plants inside an acoustic chamber, and in a greenhouse, while monitoring the plant’s
physiological parameters. We developed machine learning models that succeeded in identifying the condi-
tion of the plants, including dehydration level and injury, based solely on the emitted sounds. These informa-
tive sounds may also be detectable by other organisms. This work opens avenues for understanding plants
and their interactions with the environment and may have significant impact on agriculture.

INTRODUCTION the plant, if at all.17,21,22 Thus, the question of airborne sound

emission by plants remains unanswered.17,23,24
Plants exhibit significant changes in their phenotypes in Many animals, including herbivores and their predators,
response to stress. They differ visually, with respect to both color respond to sound.25–27 Recently, plants were also demonstrated
and shape, from unstressed plants.1–4 They also emit volatile to respond to sounds,13,28–30 e.g., by changing the expression of
organic compounds (VOCs), e.g. when exposed to drought or specific genes,29,30 or by increasing sugar concentration in their
herbivores.5,6 VOCs can also affect neighboring plants, resulting nectar.31 Thus, if plants emit airborne sounds, these sounds can
in increased resistance in these plants.7,8 Altogether, plants have potentially trigger a rapid response in nearby organisms,
been demonstrated to produce visual, chemical, and tactile including both animals and plants. Even if the emission of the
cues, which other organisms can respond to.9–12 Nevertheless, sounds is merely a result of the plant’s physiological condition,
the ability of plants to emit airborne sounds—that could poten- nearby organisms that are capable of hearing these sounds
tially be heard by other organisms—has not been sufficiently could use them for their own benefit. We therefore set to examine
explored.11,13,14 whether plants emit informative airborne sounds, which may
Plant vibrations have been described in several scenarios. serve as potential signals or cues to their environment. Here
Plants exposed to drought stress have been shown to experi- we show that plants indeed emit airborne sounds, which can
ence cavitation – a process where air bubbles form, expand be detected from several meters away, both in acoustic cham-
and collapse in the xylem, causing vibrations.15,16 These vibra- bers and in greenhouses. Moreover, we show that the emitted
tions have been recorded by connecting the recording device sounds carry information about the physiological state of the
directly to the plant.16–20 Such contact recordings do not reveal plant. By training machine learning models, we were able to
the extent to which these (and other) vibrations could result in distinguish between drought-stressed, cut, and control plants,
acoustic emissions that may be detected at a distance from based only on the sounds they emit. These results demonstrate

1328 Cell 186, 1328–1336, March 30, 2023 ª 2023 The Authors. Published by Elsevier Inc.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
 ll
Article OPEN ACCESS

Figure 1. Stressed plants emit remotely detectable ultrasounds that reveal plant condition and species
(A) Acoustic box setup. In each recording, three plants are placed inside a 50 3 100 3 150 cm3 acoustic box with two directional microphones oriented at each
plant. Using two microphones helps eliminating false detections resulting from electrical noise of the recording system and cross-plant interference.
(B) Mean number of sounds emitted during 1 h of recording by tomato and tobacco plants under two treatments: drought stress and cutting. Three control groups
were used: pot—pots with soil but without a plant; self-control—plant before treatment; and neighbor-control—untreated plants that shared the acoustic box with
treated plants. All treatment groups emitted significantly more sounds than all control groups (p < e
6, Wilcoxon test with Bonferroni correction): less than 1
detection/hour for all plant-treatment combinations in the self-control and neighbor-control groups ; 19 % n % 51 plants for each group, and no sound detection in
the pot-control group (>500 recording hours). Error bars represent standard errors.
(C) Examples of time signals of sounds emitted by: a drought-stressed tomato, a drought-stressed tobacco, a cut tomato, and a cut tobacco, normalized. Peak
dBSPL values at 10 cms, relative to 20 mPa, are noted by the arrows (see STAR Methods).

(legend continued on next page)

Cell 186, 1328–1336, March 30, 2023 1329

 ll
OPEN ACCESS Article

the potential in studying plant bioacoustics, suggest that plant 10 cm, for tomato and tobacco, respectively, and the mean
acoustic emissions may play an important role in ecology and peak frequencies (frequency with maximal energy) of these
evolution, and may have direct implications for plant monitoring sounds was 49.6 ± 0.4 kHz and 54.8 ± 1.1 kHz, respectively.
in agriculture. The mean peak intensity of the sounds emitted by cut plants
was 65.6 ± 0.2 dBSPL and 63.3 ± 0.2 dBSPL at 10.0 cm, for
RESULTS tomato and tobacco, respectively, and the mean peak fre-
quency was 57.3 ± 0.7 kHz and 57.8 ± 0.7 kHz, respectively.
To investigate plants’ airborne sound emissions, we constructed Note that due to the directionality of the microphone, this is a
a reliable recording system, where each plant is recorded simul- lower bound of the intensity (see STAR Methods).
taneously by two microphones (see Figure 1A for illustration, and Spectrograms of raw recorded sounds from drought-stressed
STAR Methods for details). First, we recorded plants within an and cut plants are shown at Figure S1A, the mean spectra of the
acoustic box and developed machine learning algorithms to sounds are shown in Figure S1B, and an example of the
classify the recorded sounds. Then we tested the system in a electronic noise is in Figure S1C. An audio sample of the actual
greenhouse, while monitoring physiological parameters of the tomato recordings, after down sampling to the audible range
recorded plants. and condensation in time, is included in Audio S1.
We trained machine learning models to classify different plant
Stressed plants emit airborne sounds that reveal their conditions and species based on their sound emissions. We
condition: Acoustic box setup divided the sounds into four groups, corresponding to the four
We recorded tomato (Solanum lycopersicum) and tobacco combinations of two plant types (tomato or tobacco), and two
(Nicotiana tabacum) plants under different treatments—drought treatments (drought stress or cutting). A binary classifier was
stress, cutting (of the stem, before recording), and controls, trained to distinguish between two equal-size groups (‘‘pair’’) in
within an acoustically isolated box. We focused on the ultrasonic each comparison (Tomato-Dry vs. Tomato-Cut; Tobacco-Dry
sound range (20–150 kHz). vs. Tobacco-Cut; Tomato-Dry vs. Tobacco-Dry; Tomato-Cut
We found that plants emit sounds, and that stressed plants— vs. Tobacco-Cut). For cross validation, the model was tested
both drought-stressed (Dry) and cut plants (Cut; see STAR only on plants that were not a part of the training process (see
Methods)—emit significantly more sounds than plants of any of STAR Methods for more details). We used a support vector ma-
the control groups (p < e
6, Wilcoxon test for each of the 12 chine (SVM) as the classifier with several methods of feature
comparisons with Holm-Bonferroni correction). Three controls extraction—basic,32,33 MFCC,34 and a scattering network.26
were used for each plant species and treatment: recording The SVM classifier with scattering network for feature extrac-
from the same plant before treatment (Self-control), recording tion achieved 70% accuracy for each of the four pairs (Figure 1F
from an untreated same-species neighbor plant (Neighbor-con- red line), significantly better than random (p < e
12 for each pair,
trol), and recording a pot with soil but without a plant (Pot; see Wilcoxon rank-sum test with Holm-Bonferroni correction). The
STAR Methods). The mean number of sounds emitted by dry same classifier was trained to discriminate between the electri-
plants was 35.4 ± 6.1 and 11.0 ± 1.4 per hour for tomato and cal noise of the system (see STAR Methods) and the sounds
tobacco, respectively, and cut tomato and tobacco plants emitted by the plants (Tomato vs. Noise, Tobacco vs. Noise)
emitted 25.2 ± 3.2 and 15.2 ± 2.6 sounds per hour, respectively and achieved above 98% accuracy for both (Figure 1F). The re-
(Figure 1B). In contrast, the mean number of sounds emitted by sults were also robust to the dimension of the descriptors and
plants from all the control groups was lower than 1 per hour. Our the scattering network specific parameters (Figure S1D). When
system did not record any sound in the Pot control (Figure 1B) using the SVM classifier with MFCC as the input features, the re-
over >500 h of recordings. sults were still significantly better than random (black line,
What does a stressed plant sound like? Figures 1C and 1D p < e
4 for each pair, corrected as above), and even when using
show examples of raw recorded time signals and their spectra only 4 ‘‘basic’’ features (see STAR Methods), the results were
as recorded from drought-stressed and cut plants, while the significantly better than random for 5 of the 6 pairs (p < e
6 for
distributions of sound peak intensity and the maximum energy each of them, adjusted; Figure 1F gray line). However, scattering
frequency of drought-stressed and cut plants are shown at Fig- network performed better than either MFCC or Basic for all the
ure 1E. The mean peak sound intensity recorded from dry pairs (p < 0.05 and p < e
6, respectively; Wilcoxon signed-
plants was 61.6 ± 0.1 dBSPL and 65.6 ± 0.4 dBSPL at rank test). Altogether, these results demonstrate that plant

(D) The normalized spectra of the sounds from (C).

(E) The recorded sounds intensity peak and the max energy frequency for the four groups—drought-stressed tomato plants, cut tomato plants, drought-stressed
tobacco plants, and cut tobacco plants.
(F) The accuracy of sound classification achieved by different feature extraction methods, with an SVM classifier. The best results were obtained using the
scattering network method for feature extraction (red line, p < e
12 for each pair). Using MFCC for feature extraction the results were also highly significant (black
dashed line, p < e
4 for each pair) and even basic methods for feature extraction allowed for better-than-random classification (gray line, p < e
6 for each pair
apart from one case: Tobacco dry vs. Tobacco cut, which was not significant with the basic method). The comparisons Tomato vs Elect. Noise and Tobacco vs
Elect. Noise refer to electrical noise of the recording system. Training set size of the two groups in each pair was equal (400 < sounds for each pair, see Table S1),
and significance levels for each pair were calculated using Wilcoxon rank-sum test with Holm–Bonferroni correction for multiple comparisons. Error bars
represent standard deviations. The classification results were reproduced using CNN models (see STAR Methods and Figure S1E).
See also Figure S1.

1330 Cell 186, 1328–1336, March 30, 2023

 ll
Article OPEN ACCESS

Figure 2. Acoustic detection of plant condi-

A Genera ng a classifier B Greenhouse experiments tion in the greenhouse
(A) Illustration of the procedure used to train a
Tomato sounds, Greenhouse noises, Recording tomato plant inside a greenhouse classifier that distinguishes between tomato sounds
recorded inside an recorded inside an and greenhouse noises. A background noises library
acous c box empty greenhouse was first generated, by recording inside an empty
greenhouse for several days. Using this library and
the library of tomato sounds recorded in an acoustic
box, we trained a convolution neural network (CNN)
classifier to distinguish between tomato sounds and
background noises.
(B) Illustration of the recordings in the greenhouse.
Tomato plants were recorded in the greenhouse.
The recorded sounds were filtered using the trained
CNN classifier, leaving only the sounds classified as
Trained Classifier tomato sounds.
Training a CNN classifier Filter-in tomato sounds (C) Confusion matrix showing the success of trained
CNN classifiers in distinguishing between tomato
sounds and background noises in a cross-validation
examination (see STAR Methods). Balanced accu-
racy score of 99.7%.
(D) Confusion matrix showing the success in dis-
tinguishing between dry and irrigated tomato plants,
Tomato Greenhouse based on 1 h of recording in the greenhouse. The
sound noise condition of the plant (dry/irrigated) was determined
Analyze the recorded tomato sounds
based on the number of recorded sounds classified
as tomato sounds: if above three, the plant was
C Signal-based classifica on D Rate-based classifica on classified as ‘‘dry’’, and otherwise as ‘‘irrigated’’.
Balanced accuracy score of 84% (p < e
5; Fisher
exact test).
See also Figure S2.

greenhouse (see ‘‘classifying sounds in

the greenhouse’’ in the STAR Methods
and Figures S2A and S2B for details).
We then trained a model on the entire
datasets (empty greenhouse noises and
dry tomato sounds from the acoustic box)
and applied it to another set of recordings
of tomatoes in the greenhouse. After
sounds carry information that can be interpreted and used for filtering out the background noises based on our model’s classi-
classification of plant type and condition. fication, the number of plant sounds per hour of recording was
highly indicative of the plant’s condition, distinguishing
Plant stress can be identified from its sounds in a drought-stressed plants from control plants with 84%
greenhouse accuracy (p < e
5, Fisher exact test; see Figures 2B and 2D
We also tested the acoustic behavior of plants in a greenhouse, and further details in Figures S2C and S2D).
in the presence of many background noises that were absent in We then tested the acoustic manifestation of the dehydration
the acoustic box (e.g. wind, air-conditioning, maintenance work). process by recording 23 tomato plants for several consecutive
In order to distinguish between the sounds generated by plants days without watering. We recorded in the greenhouse while
and background greenhouse noises, we first constructed a li- monitoring soil moisture in each pot and paired each recorded
brary of greenhouse noises, by recording in an empty green- sound with the volumetric water content (VWC) measurement
house. We then trained a convolution neural network (CNN) of the recorded plant at the time of recording. We used the
model to distinguish between these empty greenhouse noises CNN model mentioned above (Figures 2A and 2B) to separate to-
and the sounds of dry tomatoes recorded in the acoustic box mato sounds from background noises and monitored the num-
(Figure 2A; see details in STAR Methods section). The models ber of tomato-classified sounds recorded from each plant. The
obtained from this training achieved 99.7% balanced accuracy results revealed a clear temporal acoustic pattern: the plants
in a cross-validation examination (Figure 2C). Similarly, signifi- emit very few sounds when irrigated, the number of sounds
cant results were obtained when controlling for sound intensity per day increases in the following 4–5 days, and then the number
differences between plant and noise sounds and for the differ- of sounds decreases as the plant dries up (Figure 3A). The num-
ence in background sounds between the acoustic box and the ber of emitted sounds showed a bimodal pattern along the day:

Cell 186, 1328–1336, March 30, 2023 1331

 ll
OPEN ACCESS Article

one major peak in the morning hours, 08:00–12:00; and a second

smaller peak in the afternoon, 16:00–19:00 (Figure 3B). Such
‘‘midday depression’’ has previously been observed with
respect to stomatal conductance,35,36 suggesting that the two
processes may be associated. We also found a strong associa-
tion between the VWC of the soil and plant sound emission.
Consistent with our results from the acoustic box (Figure 1B),
we found that the vast majority of the plant sounds are emitted
when the VWC is <0.05, and almost no sounds are emitted
when the VWC is >0.1 (Figure 3C). Repeating the analysis
presented in Figures 3A–3C controlling for the individual varia-
tion in the number of sounds emitted by different plants resulted
in very similar results (Figures S3A–S3C).
We further examined whether the plants emitted different
sounds at different levels of dehydration. We divided the sounds
that were recorded in the experiment and classified as tomato
sounds into two groups according to the VWC of the plant at
the time of the sound emission: VWC < 0.01 and VWC > 0.05.
We then trained CNN models to distinguish between the groups
according to the sound and achieved 81% balanced accuracy in
a cross-validation examination (see STAR Methods) (Figure 3D).
Classification of sounds grouped to VWC < 0.05 and VWC > 0.05
yielded >72% balanced accuracy (Figure S3D). We conducted a
second analysis of the sounds recorded in that greenhouse
experiment, without pre-classification of the sounds to tomato
sounds and background noises. Here we defined three groups
of sounds: all sounds from the greenhouse experiment that
were associated with plant VWC lower than 0.01; all sounds
from the greenhouse experiment that were associated with plant
Figure 3. Acoustic manifestation of dehydration in tomato plants as VWC greater than 0.05; and the library of empty greenhouse
observed in long-term recordings in a greenhouse noises. We trained CNN models to distinguish between the
23 tomato plants were recorded inside a greenhouse for nine consecutive days sounds of these three groups. The results of this classification
after watering, and the volumetric water content (VWC) of their soil was measured were still highly significant (balanced accuracy >75%), even
throughout the experiment. The recorded sounds were filtered using the trained with normalized sound intensity (Figure S3E).
CNN classifier, leaving only the sounds classified as emitted by tomatoes.
To investigate the generality of our results, we also performed
(A) The number of sounds per day along nine consecutive days of dehydration.
The dots represent the average of 23 plants, while the error bars represent the
a small survey of different plant species. We successfully re-
standard errors. We find significant differences in the number of emitted corded sounds from additional plants from different taxa,
sounds between the first and second day, and between the second and third including Triticum aestivum (wheat), Zea mays (corn),
day (p values are <0.01 and <0.001, respectively; p values were calculated Vitis vinifera (Cabernet Sauvignon grapevine), Mammillaria
using Wilcoxon signed-rank tests, adjusted for 8 comparisons between pairs spinosissima (pincushion cactus) and Lamium amplexicaule
of consecutive days using Holm-Bonferroni method). The soil VWC of the (henbit) (see Figure S3F), but not from woody parts of almond
plants at the beginning of the recording was 0:21 ± 0:03 ðm3 =m3 ).
and grapevine. We thus expect that many plants emit sounds,
(B) The number of sounds per hour is plotted as function of the time of day.
Each dot represents the average number of sounds emitted in the relevant but the diversity of characteristics of these sounds are yet to
hour over all 23 plants and nine days for each plant, while the error bars be researched. We have also successfully recorded tomato
represent the standard errors. Dark gray areas represent the hours of complete plants under a different stress—infection with TMV (Figure S3F).
darkness, light gray areas show when the greenhouse lighting was on, and the We thus expect that many plants emit sounds under different
white area represents the approximate hours of natural daylight. stresses, but the diversity of characteristics of these sounds
(C) A histogram showing the total number of emitted sounds as function of the
are yet to be researched.
plant VWC during sound emission. Out of all the emitted plant sounds in this
experiment (over 20,000 sounds), 92% were emitted when VWC % 0:05, and
Finally, we investigated the correlation between the emitted
43% when VWC % 0:01. To correct for variability between plants, we sounds and the physiological state of tomato plants, focusing
repeated the analyses of (A–C) while normalizing the sounds of each plant by on the transpiration rate patterns during drought. We monitored
the overall number of sounds it emitted throughout the experiment, and found plants in the greenhouse for several consecutive days without
very similar results (see Figures S3A–S3C). watering and collected the hourly number of emitted sounds
(D) Confusion matrix showing the success of the trained CNN classifier in
as well as the hourly average transpiration rate of each plant,
distinguishing between individual sounds emitted by a plant experiencing
VWC < 0.01 and sounds emitted by a plant experiencing VWC > 0.05. This
using plant-DiTech system (see STAR Methods). We found a
classification achieved balanced accuracy score of 81% (average of leave- strong correlation between the number of sounds emitted per
one-plant-out cross validation; see STAR Methods). hour and the corresponding transpiration rate of the plant, with
See also Figure S3. a similar trend of low values during night-time, and a peak in

1332 Cell 186, 1328–1336, March 30, 2023

 ll
Article OPEN ACCESS

Figure 4. Plant sounds correlate with the

A B
transpiration rate
The sounds and transpiration rates (TR) of seven
tomato plants were monitored for several days
without watering.
(A) The number of tomato sounds per hour (black)
and the hourly TR (red) are plotted along the four
consecutive days in which each plant made the
most sounds. Each dot represents the average of
seven plants, with the bars presenting the standard
errors.
(B) Cross-correlation plot. Pearson’s R correlation
coefficient between the hourly mean number of
sounds and the hourly mean TR is plotted (y axis) as
a function of the time lag (in hours) between the timing of the sounds and the TR (x axis). Each dot represents the average correlation across all seven plants for the
lag specified in the x axis, while the bars represent the standard errors.
See also Figure S4.

the late morning-noon time. Nevertheless, we see that the daily through all the trachea in the cut stem. In accordance, sounds
transpiration rate decreases as dehydration continues, while were emitted by cut plants for a shorter period of time than by
the number of emitted sounds does not (Figure 4). dry plants. Of the sounds emitted by plants that were both cut
and dry, the majority were classified as ‘‘cut’’ in the first day,
DISCUSSION but the picture is reversed in the following days, with a majority
of sounds classified as ‘‘dry’’ (Figure S4C). (3) A 3D acoustic
Our results demonstrate that plants emit remotely detectable simulation shows that sounds emitted from the trachea would
and informative airborne sounds under stress (Figure 1). The radiate from the stem in all directions (Figure S4D). This is consis-
plant emissions that we report, in the ultrasonic range of 20– tent with the results of our two-microphone recording system,
100 kHz, could be detected from a distance of 3–5 m, by many which picked up sound on two sides of the stem they were
mammals and insects (given their hearing sensitivity, e.g., directed to (Figure 1A). (4) The frequency range of cavitation-
mice37 and moths27,38; see STAR Methods for details on the esti- related vibrations partially overlaps with the sounds we re-
mation of animal detection ranges). We succeeded in differenti- corded.17 When recorded with contact sensors, cavitation is
ating between sounds emitted under two different stress condi- usually also characterized by additional higher frequencies that
tions—dry and cut plants—with an accuracy of 70% using are beyond the sensitivity range of our microphones and that
supervised machine learning methods (Figure 1), and distin- would attenuate rapidly in air. Yet, only the vibrations that result
guishing between drought-stressed and control plants in a in airborne sounds (which we report here) are the ones that have
greenhouse, based only on the sounds they emit (accuracy of a potential of affecting other organisms and human-sensors that
84%, Figures 2B and 2D). We monitored the distribution of plant are not in direct contact with the plant. When considering groups
sounds with respect to days of dehydration, time of day, of plants, the advantage of tracking airborne sound may be even
soil moisture (Figure 3), and transpiration rate (Figure 4). Our re- greater, as a single artificial sensor for airborne sounds could
cordings revealed that the hourly pattern of sound emission detect sounds requiring multiple contact sensors.
correlates with the plant’s transpiration rate, while the daily num- A potential application of our results can be for monitoring
ber of sounds is a hump-shaped function of plant dehydration, plants in the field or greenhouse.40 Specifically, plant sound
increasing during the first days of dehydration, and declining emissions could offer a way for monitoring crops water and
as the plant dries up. The sounds emitted by the plants at high possibly disease states—questions of crucial importance in agri-
and low levels of dehydration are different, and we succeeded culture.41 More precise irrigation can save up to 50% of the water
classifying them with an accuracy of 81% (Figure 3). These find- expenditure and increase the yield, with dramatic economic im-
ings can alter the way we think about the plant kingdom, which plications.41,42 In times when more and more areas are exposed
has been considered to be almost silent until now.23 to drought due to climate change,43 efficient water use becomes
One potential mechanism that may be responsible for the even more critical, for both food security and ecology. Our re-
emission of at least part of the sounds we record is cavitation sults, demonstrating the ability to distinguish between drought-
in the stem.16 Several findings support this: (1) we found that stressed and control plants based on plant airborne sounds,
the frequencies of the sounds emitted by different plant species open an avenue of research in the field of precision agriculture.
correspond to their trachea diameter, with wider tracheas in the We have shown that plant sounds can be effectively classified
plants emitting lower sounds (Figures S4A and S4B), consistent by machine learning algorithms. We thus suggest that other or-
with the observed negative association between xylem diameter ganisms may have evolved to classify these sounds as well
and resonance frequency.39 (2) The different sounds emitted un- and respond to them. For instance, many moths—some of
der drying and cutting (Figure 1) are in accordance with the them using tomato and tobacco as hosts for their larvae44,45—
different gas dynamics of the plant in these two processes: while can hear and react to ultrasound in the frequencies and inten-
drying is gradual, with a low rate of air-seeding and reduced sities that we recorded.25–27 Nearby plants may also respond
pressure, cutting involves a rapid and significant air-seeding to the sounds emitted by plants. Plants were already shown to

Cell 186, 1328–1336, March 30, 2023 1333

 ll
OPEN ACCESS Article

react to sounds13,28–31,46 and specifically to increase their SUPPLEMENTAL INFORMATION

drought tolerance in response to sounds.47,48 Could plants
Supplemental information can be found online at https://doi.org/10.1016/j.cell.
potentially respond adaptively to the sounds of their drought-
2023.03.009.
stressed or injured neighbors? We suggest that more investiga-
tion in the plant bioacoustics field, and particularly in the ability of ACKNOWLEDGMENTS
plants to emit and react to sounds under different conditions and
environments, may reveal a pathway of signaling between plants We thank Daniel Chamovitz, Gal Chechik, Tuvik Beker, Marcus W. Feldman,
and their environment. Judith Berman, and Ella Sklan for comments on the paper; Guido Sessa,
Doron Teper, Guy Sobol, Yura Pupov, Rotem Shteinshleifer, Odelia Pisanty,
Eilon Shani, Mor Binder, Meirav Leibman-Markus, and Alin Finkelstein for help-
Limitations of the study ing with plants materials; Aviv Dombrovsky for TMV; Yoel Shkolnisky, Marine
Although our study demonstrates that plant emit informative Veits, Ilia Raysin, Uri Obolski, Yoav Ram, Eyal Zinger, Yael Gurevich, Eylon Ta-
airborne sounds under stress, there are a few open issues: mir, Yuval Sapir, Yaara Blogovski, and Ruth Cohen-Khait for comments on the
First, our results were obtained on a limited number of plant way; Yonatan Bendett for comments and artwork. The Titan Xp used for this
species, and should be tested on additional species of plants research was donated by the NVIDIA Corporation.
Funding: The research has been supported in part by Bikura 2308/16 (L.H.,
from different families. In a small survey, we successfully re-
Y.Y.), Bikura 2658/18 (L.H., Y.Y.), ISF 2064/18 (L.H.), Tel Aviv University Center
corded sounds from plants of 5 additional taxa (see Figure S3F).
for AI and Data Science (TAD) (L.H., Y.Y.), by the Manna Center Program for
We thus expect that many plants emit sounds, but the diversity Food Safety and Security fellowships (I.K.), and by the Clore Foundation
of characteristics of these sounds are yet to be identified. Sec- Scholars Program (O.L.-E.).
ond, future studies could explore the sounds emitted under
different conditions. We observed sound emission in plants AUTHOR CONTRIBUTIONS
exposed to drought, cutting, or TMV infection (Figures 1 and
S3F). Other potential conditions include different pathogens, L.H. and I.K. conceived the study. L.H., Y.Y., I.K., O.L.-E. and N.S. designed
the research. I.K., O.L.-E., K.S., R.G., Y.Z., Y.A., S.N., R.P., and G.S. per-
cold, herbivores attack, UV radiation, and other life stages of
formed the experiments. O.L.-E., R.S., I.K., A.B., Y.A., C.A., and L.H. analyzed
the plant species, such as flowering. Third, our understanding the data. L.H., Y.Y., and N.S. supervised the experiments. L.H. and Y.Y. super-
of the sound emission mechanism is still rudimentary. This is vised the analysis. I.K. and O.L.-E. contributed equally to the study. L.H. and
an area for future investigation. Finally, our results were Y.Y. contributed equally to the study. All authors discussed the results and
obtained in either a controlled acoustic environment (an acous- took part in writing the manuscript.
tic chamber) or a semi-natural environment (greenhouse).
Recording and analysis of plant sounds in the field, with a wider DECLARATION OF INTERESTS

range of background noises, would present additional

Itzhak Khait, Raz Sharon, Yossi Yovel, and Lilach Hadany submitted a patent
challenges. application related to this work: Plant-monitor. Patent application number
WO2020039434A1. (2020). The remaining authors declare no competing
STAR+METHODS interests.

Received: July 27, 2021

Detailed methods are provided in the online version of this paper
Revised: November 29, 2022
and include the following:
Accepted: March 6, 2023
Published: March 30, 2023
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY REFERENCES
B Lead contact
B Data and code availability 1. Hsiao, T.C. (1973). Plant responses to water stress. Annu. Rev. Plant Phys-
d EXPERIMENTAL MODEL AND SUBJECT DETAILS iol. 24, 519–570.
d METHOD DETAILS 2. Jackson, R.D. (1986). Remote sensing of biotic and abiotic plant stress.
B Recording protocol Annu. Rev. Phytopathol. 24, 265–287.
B Data pre-processing 3. Mahlein, A.-K. (2016). Plant disease detection by imaging sensors–
B Acoustic-box drought stress experiment parallels and specific demands for precision agriculture and plant pheno-
typing. Plant Dis. 100, 241–251.
B Acoustic-box cut stress experiment
B Greenhouse drought experiment 4. Potters, G., Pasternak, T.P., Guisez, Y., Palme, K.J., and Jansen, M.A.K.
(2007). Stress-induced morphogenic responses: growing out of trouble?
B Greenhouse dehydration experiment
Trends Plant Sci. 12, 98–105.
B Greenhouse whole plant transpiration rate
5. Holopainen, J.K., and Gershenzon, J. (2010). Multiple stress factors and
B Classifying sounds recorded in the acoustic box
the emission of plant VOCs. Trends Plant Sci. 15, 176–184.
B Convolution neural networks (CNN)
6. Paré, P.W., and Tumlinson, J.H. (1999). Plant volatiles as a defense against
B Leave-one-plant-out cross validation (LOPO-CV) insect herbivores. Plant Physiol. 121, 325–332.
B Classifying sounds recorded in the greenhouse
7. Heil, M., and Karban, R. (2010). Explaining evolution of plant communica-
B Estimation of the animal detection range tion by airborne signals. Trends Ecol. Evol. 25, 137–144.
B Microphone calibration 8. Dolch, R., and Tscharntke, T. (2000). Defoliation of alders (Alnus glutinosa)
B Estimating the acoustic box properties affects herbivory by leaf beetles on undamaged neighbours. Oecologia
d QUANTIFICATION AND STATISTICAL ANALYSIS 125, 504–511.

1334 Cell 186, 1328–1336, March 30, 2023

 ll
Article OPEN ACCESS

9. Karban, R. (2008). Plant behaviour and communication. Ecol. Lett. 11, 32. Acevedo, M.A., Corrada-Bravo, C.J., Corrada-Bravo, H., Villanueva-Riv-
727–739. era, L.J., and Aide, T.M. (2009). Automated classification of bird and
10. Falik, O., Mordoch, Y., Quansah, L., Fait, A., and Novoplansky, A. (2011). amphibian calls using machine learning: A comparison of methods.
Rumor has it.: relay communication of stress cues in plants. PLoS One 6, Ecol. Inf. 4, 206–214.
e23625. https://doi.org/10.1371/journal.pone.0023625. 33. Giannakopoulos, T., and Pikrakis, A. (2014). Introduction to Audio Anal-
11. Chamovitz, D. (2012). What a Plant Knows: A Field Guide to the Senses of ysis: A MATLAB Approach (Academic Press).
Your Garden-And beyond (Scientific American/Farrar, Straus and Giroux). 34. Ellis, D.P. (2005). {PLP} and {RASTA}(and {MFCC}, and inversion) in
12. Lev-Yadun, S. (2016). Defensive (Anti-herbivory) Coloration in Land Plants {M} atlab.
(Springer). 35. Brodribb, T.J., and Holbrook, N.M. (2004). Diurnal depression of leaf hy-
13. Hassanien, R.H., HOU, T.-z., LI, Y.-f., and LI, B.-m. (2014). Advances in ef- draulic conductance in a tropical tree species. Plant Cell Environ. 27,
fects of sound waves on plants. J. Integr. Agric. 13, 335–348. 820–827.
14. Gagliano, M., Mancuso, S., and Robert, D. (2012). Towards understanding 36. Gosa, S.C., Koch, A., Shenhar, I., Hirschberg, J., Zamir, D., and Moshe-
plant bioacoustics. Trends Plant Sci. 17, 323–325. lion, M. (2022). The potential of dynamic physiological traits in young to-
mato plants to predict field-yield performance. Plant Sci. 315, 111122.
15. Tyree, M.T., and Sperry, J.S. (1989). Vulnerability of xylem to cavitation
and embolism. Annu. Rev. Plant Physiol. Plant Mol. Biol. 40, 19–36. 37. Heffner, H.E., and Heffner, R.S. (1985). Hearing in two cricetid rodents:
16. Cochard, H., Badel, E., Herbette, S., Delzon, S., Choat, B., and Jansen, S. Wood rat (Neotoma floridana) and grasshopper mouse (Onychomys leu-
(2013). Methods for measuring plant vulnerability to cavitation: a critical re- cogaster). J. Comp. Psychol. 99, 275–288.
view. J. Exp. Bot. 64, 4779–4791. 38. Moir, H.M., Jackson, J.C., and Windmill, J.F.C. (2013). Extremely high fre-
17. De Roo, L., Vergeynst, L., De Baerdemaeker, N., and Steppe, K. (2016). quency sensitivity in a ‘simple’ear. Biol. Lett. 9, 20130241.
Acoustic emissions to measure drought-induced cavitation in plants. 39. Dutta, S., Chen, Z., Kaiser, E., Matamoros, P.M., Steeneken, P.G., and
Appl. Sci. 6, 71. Verbiest, G.J. (2022). Ultrasound Pulse Emission Spectroscopy Method
18. Zweifel, R., and Zeugin, F. (2008). Ultrasonic acoustic emissions in to Characterize Xylem Conduits in Plant Stems. Research 2022.
drought-stressed trees – more than signals from cavitation? New Phytol. 40. Khait, I., Sharon, R., Yovel, Y., and Hadany, L. (2020). Plant-monitor.
179, 1070–1079. https://doi.org/10.1111/j.1469-8137.2008.02521.x. Patent Application Number WO2020039434A1. patent application
19. Ikeda, T., and Ohtsu, M. (1992). Detection of xylem cavitation in field-- WO2020039434A1.
grown pine trees using the acoustic emission technique. Ecol. Res. 7, 41. Playán, E., and Mateos, L. (2006). Modernization and optimization of irriga-
391–395. tion systems to increase water productivity. Agric. Water Manag. 80,
20. Jackson, G., and Grace, J. (1996). Field measurements of xylem cavita- 100–116.
tion: are acoustic emissions useful? J. Exp. Bot. 47, 1643–1650. 42. Sadler, E., Evans, R., Stone, K., and Camp, C. (2005). Opportunities for
21. Bailey, N.W., Fowler-Finn, K.D., Rebar, D., and Rodrı́guez, R.L. (2013). conservation with precision irrigation. J. Soil Water Conserv. 60, 371–378.
Green symphonies or wind in the willows? Testing acoustic communica- 43. Allen, C.D., and Breshears, D.D. (1998). Drought-induced shift of a forest–
tion in plants. Behav. Ecol. 24, 797–798. ars228. woodland ecotone: rapid landscape response to climate variation. Proc.
22. ten Cate, C. (2013). Acoustic communication in plants: do the woods really Natl. Acad. Sci. USA 95, 14839–14842.
sing? Behav. Ecol. 24, 799–800. 44. Specht, A., de Paula-Moraes, S.V., and Sosa-Gómez, D.R. (2015). Host
23. Gagliano, M. (2012). Green symphonies: a call for studies on acoustic plants of Chrysodeixis includens (Walker)(Lepidoptera, Noctuidae, Plusii-
communication in plants. Behav. Ecol. 24, 789–796. ars206. nae). Rev. Bras. Entomol. 59, 343–345.
24. Jung, J., Kim, S.-K., Kim, J.Y., Jeong, M.-J., and Ryu, C.-M. (2018). 45. Liu, Z., Li, D., Gong, P., and Wu, K. (2004). Life table studies of the cotton
Beyond chemical triggers: evidence for sound-evoked physiological reac- bollworm, Helicoverpa armigera (Hübner)(Lepidoptera: Noctuidae), on
tions in plants. Front. Plant Sci. 9, 25. different host plants. Environ. Entomol. 33, 1570–1576.
25. Miller, L.A., and Surlykke, A. (2001). How Some Insects Detect and Avoid Be- 46. Rodrigo-Moreno, A., Bazihizina, N., Azzarello, E., Masi, E., Tran, D., Bou-
ing Eaten by Bats: Tactics and Countertactics of Prey and Predator Evolu- teau, F., Baluska, F., and Mancuso, S. (2017). Root phonotropism: Early
tionarily speaking, insects have responded to selective pressure from bats signalling events following sound perception in Arabidopsis roots. Plant
with new evasive mechanisms, and these very responses in turn put pressure Sci. 264, 9–15.
on bats to ‘‘improve’’ their tactics. Bioscience 51, 570–581. 47. López-Ribera, I., and Vicient, C.M. (2017). Drought tolerance induced by
26. Spangler, H.G. (1988). Moth hearing, defense, and communication. Annu. sound in Arabidopsis plants. Plant Signal. Behav. 12, e1368938. https://
Rev. Entomol. 33, 59–81. doi.org/10.1080/15592324.2017.1368938.
27. Fullard, J.H., Dawson, J.W., and Jacobs, D.S. (2003). Auditory encoding 48. Jeong, M.-J., Cho, J.-I., Park, S.-H., Kim, K.-H., Lee, S.K., Kwon, T.-R.,
during the last moment of a moth’s life. J. Exp. Biol. 206, 281–294. Park, S.-C., and Siddiqui, Z.S. (2014). Sound frequencies induce drought
28. Mishra, R.C., Ghosh, R., and Bae, H. (2016). Plant acoustics: in the search tolerance in rice plant. Pak J Bot 46, 2015–2020.
of a sound mechanism for sound signaling in plants. J. Exp. Bot. 67, 49. Gibly, A., Bonshtien, A., Balaji, V., Debbie, P., Martin, G.B., and Sessa, G.
4483–4494. (2004). Identification and expression profiling of tomato genes differen-
29. Jeong, M.-J., Shim, C.-K., Lee, J.-O., Kwon, H.-B., Kim, Y.-H., Lee, S.-K., tially regulated during a resistance response to Xanthomonas campestris
Byun, M.-O., and Park, S.-C. (2008). Plant gene responses to frequency- pv. Mol. Plant Microbe Interact. 17, 1212–1222.
specific sound signals. Mol. Breeding 21, 217–226. 50. Gera, A., Loebenstein, G., and Shabtai, S. (1983). Enhanced tobacco
30. Ghosh, R., Mishra, R.C., Choi, B., Kwon, Y.S., Bae, D.W., Park, S.-C., mosaic virus production and suppressed synthesis of a virus inhibitor in
Jeong, M.-J., and Bae, H. (2016). Exposure to sound vibrations lead to protoplasts exposed to antibiotics. Virology 127, 475–478.
transcriptomic, proteomic and hormonal changes in Arabidopsis. Sci. 51. Sessa, G., Yang, X.-Q., Raz, V., Eyal, Y., and Fluhr, R. (1995). Dark induc-
Rep. 6, 33370–33417. tion and subcellular localization of the pathogenesis-related PRB-1b pro-
31. Veits, M., Khait, I., Obolski, U., Zinger, E., Boonman, A., Goldshtein, A., Sa- tein. Plant Mol. Biol. 28, 537–547.
ban, K., Seltzer, R., Ben-Dor, U., Estlein, P., et al. (2019). Flowers respond 52. Zinger, E., Gueijman, A., Obolski, U., Ram, Y., Ruby, E., Binder, M., Ye-
to pollinator sound within minutes by increasing nectar sugar concentra- chieli, N., Ohad, N., and Hadany, L. (2019). Less fit Lamium amplexicaule
tion. Ecol. Lett. 22, 1483–1492. plants produce more dispersible seeds. Sci. Rep. 9, 6299.

Cell 186, 1328–1336, March 30, 2023 1335

 ll
OPEN ACCESS Article
53. Halperin, O., Gebremedhin, A., Wallach, R., and Moshelion, M. (2017). High- 56. Andén, J., and Mallat, S. (2014). Deep scattering spectrum. IEEE Trans.
-throughput physiological phenotyping and screening system for the char- Signal Process. 62, 4114–4128.
acterization of plant–environment interactions. Plant J. 89, 839–850.
57. Sifre, L., Kapoko, M., Oyallon, E., and Lostanlen, V. (2013). Scatnet: A
54. Dalal, A., Bourstein, R., Haish, N., Shenhar, I., Wallach, R., and Moshelion, MATLAB Toolbox for Scattering Networks.
M. (2019). Dynamic physiological Phenotyping of drought-stressed pep-
per plants treated with ‘‘productivity-enhancing’’ and ‘‘survivability- 58. Krivoruchko, K., Goldshtein, A., Boonman, A., Eitan, O., Ben-Simon, J.,
enhancing’’ biostimulants. Front. Plant Sci. 10, 905. Thong, V.D., and Yovel, Y. (2021). Fireflies produce ultrasonic clicks
during flight as a potential aposematic anti-bat signal. iScience 24,
55. Dalal, A., Shenhar, I., Bourstein, R., Mayo, A., Grunwald, Y., Averbuch, N.,
102194.
Attia, Z., Wallach, R., and Moshelion, M. (2020). A telemetric, gravimetric
platform for real-time physiological phenotyping of plant–environment in- 59. Schneider, C.A., Rasband, W.S., and Eliceiri, K.W. (2012). NIH Image to
teractions. JoVE, e61280. ImageJ: 25 years of image analysis. Nat. Methods 9, 671–675.

1336 Cell 186, 1328–1336, March 30, 2023

 ll
Article OPEN ACCESS

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Bacterial and virus strains
Tobacco Mosaic Virus (TMV) Gera et al.50 N/A
Deposited data
Recordings of plant sounds This paper https://doi.org/10.5061/dryad.jwstqjqf7
Experimental models: Organisms/strains
Solanum lycopersicum ‘Hawaii 7981’ Gilby et al.49 N/A
Nicotiana tabacum ‘Samsun NN’ Sessa et al.51 N/A
Triticum aestivum (cv fielder) ICCI N/A
Zea mays (cv B73) ICCI N/A
Vitis vinifera)cv Cabernet Sauvignon) Sdot Hemed nursery N/A
Mammilaria spinosissima TAU Botanical Garden N/A
Lamium amplexicaule Zinger et al.52 N/A
Software and algorithms
Python version 3.6 Python Software Foundation https://www.python.org
MATLAB 8.3 The MathWork Inc. https://www.mathworks.com/
ImageJ Schneider et al.59 https://imagej.nih.gov/ij/download.html
Comsol Multiphysics simulation software COMSOL https://www.comsol.com/
MFCC rastamat https://www.ee.columbia.edu/
dpwe/resources/matlab/rastamat/
Scattering networks scatnet-0.2 https://www.di.ens.fr/data/software/scatnet/
LIBSVM: A library for support vector machines libsvm-3.21 http://www.csie.ntu.edu.tw/cjlin/libsvm
CNN sound classifier This paper https://doi.org/10.5281/zenodo.7612742
Other
Condenser ultrasound microphones CM16 Avisoft Bioacoustics https://www.avisoft.com/
ultrasound-microphones/cm16-cmpa/
UltraSoundGate 1216H A/D converter Avisoft Bioacoustics https://www.avisoft.com/ultrasoundgate/1216h/

RESOURCE AVAILABILITY

Lead contact
Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Lilach Hadany (lilach.
[email protected]).

Data and code availability

d Code for the CNN binary classifier training function can be found on Zenodo: https://doi.org/10.5281/zenodo.7612742.
d Data including recordings of plants under different conditions can be found on Dryad: https://doi.org/10.5061/dryad.jwstqjqf7.
d Additional information can be obtained from the lead contact.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

S. lycopersicum ‘Hawaii 7981’.49 The plants for the experiments presented in Figures 1, 2, and 3 were grown in a growth room at 25 C
and kept in long-day conditions (16 h day, 8 h night). The plants used in the transpiration experiment (Figure 4) were grown in a green-
house. The plants used in the infection experiment (Figure S3F) were grown in a greenhouse and infected with TMV.50 The plants were
tested in the experiments 5–8 weeks after germination.
N. tabacum ‘Samsun NN’.51 The plants were grown in a growth room at 25 C and kept in long-day conditions (16 h day, 8 h night).
T. aestivum (cv fielder), grown in a greenhouse.

Cell 186, 1328–1336.e1–e4, March 30, 2023 e1

 ll
OPEN ACCESS Article

Z. mays (cv B73), grown in a greenhouse.

Vitis vinifera (cv Cabernet Sauvignon) grown in a greenhouse.
Mammilaria spinosissima, grown in a greenhouse.
L. amplexicaule,52 grown in a growth room, long-day conditions.

METHOD DETAILS

Recording protocol
In the acoustic box
The recordings were performed in a 50 3 100 3 150 cm3 acoustically isolated custom-made 1.7cm thick wooden box tiled with 6cm
acoustic foam on all sides to minimize echoes. The walls, floor and ceiling of the acoustic box had a maximum Target Strength of
40
dB at 20–30 kHz, dropping off to
50 dB at 65–70 kHz, and the box was located in the quiet basement of the faculty of life sciences
TAU, behind thick walls. The box’s walls dampened sounds generate outside the box by at least 100 dB SPL in the relevant frequency
range (see below details regarding the estimation of the acoustic box acoustics). Two cable holes, 2 cm radius each, were located in
two corners of the box and covered with PVC and acoustic foam. Inside the acoustic box were only the recorded plants and 6 mi-
crophones connected to an UltraSoundGate 1216H A/D converter (Avisoft). The PC and all the electricity connections were in the
room outside the acoustic box. Two USB cables connected the PC to the 1216H device inside the box, through the holes. There
was no light inside the acoustic box, and the testing was performed in the dark.
The recordings were performed using a condenser CM16 ultrasound microphone (Avisoft), digitized using an UltraSoundGate
1216H A/D converter (Avisoft), and stored onto a PC. The sampling rate was 500 kHz per channel. We filtered the recordings using
20 kHz high-pass filter. A recording was saved only if triggered with a sound which exceeded 2% of the maximum dynamic range of
the microphone. In such cases, a recording of length 1.5 s was saved, starting half a second before the trigger and lasting 1 s after the
trigger. Each recording produced a 6-channel wav file (a channel for each microphone). Two microphones were directed at each
plant stem, keeping a clear line of sight with no dampening leaves between the stem and the microphones, from a distance of
10 cm. Only sounds that were recorded by both microphones directed at the same plant were considered as ‘‘plant sounds’’ in
the analysis afterward. The frequency responses of the microphones can be found in Avisoft website. No sounds were recorded
by both microphones from empty pots (with either dry or wet soil), over several weeks of recordings, demonstrating the effectiveness
of the box. No sounds were recorded from empty pots even when sharing a box with stressed plants, suggesting that cross talk
between microphones was not significant in our setting.
In the greenhouse
The recordings were performed in a greenhouse in Tel-Aviv University. During the greenhouse experiment, the only plants that were
present inside the greenhouse were the recorded plants. The recordings were performed using the same hardware and setting, as
mentioned in the acoustic box recording section, with three plants in each round and two microphones oriented at each plant, apart
from the acoustic box itself.

Data pre-processing
Data processing was performed offline using python 3.6 code we developed, with the following steps: 1. Identifying the microphone
that had recorded the highest intensity peak in each 1.5 s recording file. 2. Selecting only the sounds that were detected by two
microphones oriented at the same plant at the same time, and saving the one that triggered the recording for further analysis. 3.
Focusing on a short segment of 2ms around the peak. A processed recording includes a segment of 1ms before and 1ms after
the peak of a recorded sound that triggered the system to record. ‘‘Greenhouse noise’’ sounds were obtained when the greenhouse
included only acoustic equipment without plants or pots, by the two microphones (thus excluding ‘‘Electrical noise’’ registered on a
single microphone).

Acoustic-box drought stress experiment

Each plant was recorded twice: first before drought treatment (‘‘self-control’’), and again after it. In the first recording, all the plants
were healthy and their soil was moist. Then, for 4–6 days, half of the plants were watered while the other half were not, until the soil
moisture in the pots of un-watered plants decreased below 5%. Then, the plants were recorded again at the same order. In each
recording session three plants were recorded simultaneously for 1 h and each triplet of plants included at least one watered and
one un-watered plant to allow ‘‘neighbors-control’’ – watered plants that were recorded while sharing the acoustic box with un-wa-
tered plants. Soil moisture content was recorded using a hand-held digital soil moisture meter - Lutron PMS-714.

Acoustic-box cut stress experiment

The experiment followed the experimental design of the drought stress experiment described above, but drought stress was re-
placed with cutting of the plant. Here the pot soil was kept moist for all the plants throughout the experiment. The plants included
in the treatment group were cut with scissors close to the ground right before the recording started. The severed part of the plant,
disconnected from the roots, was recorded. We used the same controls of the drought stress experiment.

e2 Cell 186, 1328–1336.e1–e4, March 30, 2023

 ll
Article OPEN ACCESS

Greenhouse drought experiment

In each recording session three plants were recorded simultaneously for 1 h. All the recorded plants were grown in a growth room,
and were brought to the greenhouse only for the recording session. Each plant was recorded either one day after irrigation (control
plants) or 4–6 days after irrigation (drought-stressed plants). The results of that experiment are presented in Figure 2.

Greenhouse dehydration experiment

In each recording session three plants were recorded simultaneously for several consecutive days without watering. All the recorded
plants were grown in a growth room, and were brought to the greenhouse only for the recording session. In addition, the soil of each plant
was monitored every 30 min throughout the experiment, using Decagon GS3 sensors. The results of that experiment are presented in
Figure 3.

Greenhouse whole plant transpiration rate

Whole-plant physiological performance was monitored with the functional phenotyping system Plant-Ditech platform.53 The Plant-
Ditech system was calibrated before the experiment start, and 5–7 weeks old plants were used for measurements (same as for the
drought and dehydration experiments). Plants were measured along several consecutive days without watering. The transpiration
rate (TR) of the plants during the course of the experiment were determined using standard previously described protocols.53–55
The results of that experiment are presented in Figure 4.

Classifying sounds recorded in the acoustic box

Our classification method was composed of two main stages. First, we extracted various acoustic features from the raw recorded
signals (see data Pre-processing section). Second, we trained a model to classify plant sounds into classes based on the feature
representation obtained in the first stage. A separate model was trained for each comparison. We used three methods of feature
extraction: (a) Deep scattering Network, as described in Andén and Mallat56 (red dotted line in Figure 1F). This method extends
MFCC (Mel-frequency cepstral coefficients) while minimizing information loss. We used the implementation by ScatNet,57 with Mor-
let wavelets. The results were robust to the dimension of descriptors and the scattering network specific parameters: number of
layers used; time support of low pass filter; and Q-Factor (Figure S1D). The values of the specific parameters used in this work
are shown in Table S2. (b) MFCC feature extraction (dashed black line in Figure 2B). We used the Dan Ellis implementation.34 (c) Basic
features. The basic features we used were energy, energy entropy, spectral entropy, and maximum frequency (gray line in Fig-
ure 1F).32,33 We used SVM with Radial kernel with the LIBSVM implementation as classifier. After feature extraction, we used Z score
for normalization of the features and PCA to reduce the dimensionality of the problem. We conducted LOPO-CV examination for
evaluation of model accuracy (see below).
We repeated the analysis using binary CNN models (see below), over either raw sound vectors or sound vectors normalized for
intensity (by dividing each value by the vector’s maximal value), with similar levels of accuracy (Figure S1E). The numbers of plants
in each group are shown at the Table S3.

Convolution neural networks (CNN)

For the implementation of convolution neural networks we used Python 3.6 with keras package and tensorflow backend. All of the
CNN models presented in this manuscript (except for the analysis in Figure S3E) are based on the same network architecture. The
network is composed of three blocks, each including two 1D convolution layers with a ‘relu’ activation, followed by a maxpooling
layer and a dropout layer. These three blocks are followed by one fully-connected layer with ‘relu’ activation and another dropout
layer. At the end, we add a fully-connected layer of size 1 with sigmoid activation. The model is trained for 50 epochs with binary
‘crossentropy’ loss function and ’adam’ optimizer.
For the analysis presented in Figure S3E (3-categories classification) we replace the last layer (of size 1 with sigmoid activation) by a
fully-connected layer of size 3, with ‘softmax’ activation. This model is trained with ‘categorical_crossentropy’ loss function.

Leave-one-plant-out cross validation (LOPO-CV)

In order to obtain robust evaluations of the models we trained, and to verify that individual differences between plants do not drive the
results, we conducted leave-one-plant-out cross validation (LOPO-CV) examinations throughout the paper. In each step of the CV, we
left all of the emitted sounds of one plant for testing, and used the rest of the sounds for training. We repeated this process so that each
plant was used for testing exactly once. This, the number of CV steps is equal to the number of plants in the analysis. For each CV, we
then constructed a confusion matrix by summing all of the accurate and wrong predictions. We sum all these matrices to obtain one
confusion matrix that summarizes the success from the entire CV process, from which we derived the balanced accuracy score.
When one of the examined groups is of noises (either electric noises or greenhouse noises), where the sounds are not emitted by
plants, we randomly split the sounds into several groups that will be used for the CV. For examination of the SVM models, in each CV
step we constructed a training set, including two groups of equal size, by randomly omitting samples from the larger group.

Cell 186, 1328–1336.e1–e4, March 30, 2023 e3

 ll
OPEN ACCESS Article

Classifying sounds recorded in the greenhouse

For classification of greenhouse noises and drought-stressed tomato sounds, we used a convolution neural network (see CNN
section above). We first examined the performance of the CNN models on our sound libraries: dehydrated tomato sounds from
the acoustic box and empty greenhouse noises, applying LOPO-CV. This analysis yielded a balanced accuracy score of 99:7%.
We repeated the evaluation analysis with two forms of normalization: (a) for sound intensity (by dividing each value in the sound
vector by the vector’s maximal absolute value); (b) for background noise, by superimposing each sound vector obtained from the
acoustic box on a randomly chosen background sound from the greenhouse and each sound from the greenhouse on a randomly
chosen background sound from the acoustic box (see Figure S2B for illustration). The results were very similar and highly significant
after these two corrections.
For the application of the CNN model on sounds recorded in the greenhouse experiments (including Figures 2D, 3, 4), classifying
each sound as either a tomato sound or a greenhouse noise, we trained a CNN model over the entire libraries of dry tomato sounds
(from the acoustic box) and empty greenhouse noises.

Estimation of the animal detection range

The distance at which a moth/mouse would detect the clicks produced by a tomato plant was calculated using the sound propaga-
tion equation: cSPL – 20*log10(R/R1) – a(R-R1) + log2.7 = HT, where HT is hearing threshold of a moth or a grasshopper mouse (set to
35 and 10 dB SPL Re 20 mPa respectively37,38), cSPL is the mean emitted plant sound pressure level set to 65 dB SPL based on our
recordings. R is the detection distance (m), R1 is the recording distance of the plant and a is a frequency dependent atmospheric
attenuation factor (1.6 dB/m at the most intense emission frequency) for the average temperature and humidity levels in the area
(26 C at night and 80% respectively). Note that some researchers argue that HT should be 10 or even 0 dB SPL Re 20 mPa in which
case the detection range would further increase, but we preferred an underestimate to be on the safe side.

Microphone calibration
The Avisoft CM16 microphone’s sensitivity is 500 mV/Pa and is almost flat between 10 and 120 kHz (see http://www.avisoft.com/
ultrasound-microphones/cm16-cmpa/). The sensitivity of the microphones used in this experiment were calibrated using a calibrated
GRAS (40DP) microphone by emitting signals covering the relevant frequency range toward the two microphones when positioned at
the same distance relative to the speaker. For playback, an ultrasonic speaker (Vifa, connected to an Avisoft 116 D/A converted) was
used to play back sound sweeps in the relevant frequency range (1 ms tones at frequencies between 70 and 20kHz(. This procedure
is routinely used in the lab for various acoustic measurements (see e.g.,58). Our GRAS microphone is routinely calibrated using a Lar-
son Davis CAL150 calibrator with a 1kHz 94 dB SPL calibration tone. The CM16 microphone is much more directional than the GRAS
so its calibration is accurate only when the sound is arriving on-axis. CM16 directionality can be found here (http://www.avisoft.com/
ultrasound-microphones/cm16-cmpa/). Our sound intensity estimates are thus a lower bound on the actual intensity.

Estimating the acoustic box properties

The echoic properties of the chamber were measured by playing a 70 to 20 kHz pulse of 1 ms duration through a Vifa loudspeaker with
a loudness of 116 dB SPL at 30 kHz. The foam at a distance of 93 cm from the speaker received 97 dB SPL, resulting in strongest
echoes of 57 dB SPL at 30 kHz with virtually no energy visible above 40 kHz. Absolute sound pressure levels from the loudspeaker
were measured by a GRAS instrumentation microphone 40DP connected to an Hm116 Avisoft sampling board, sampling at 375 kHz.
Echoes were measured with same CM16 microphone used in the plant data acquisition. The GRAS instrumentation microphone was
calibrated during the measurements using a Larson Davis CAL150 calibrator with a 1kHz 94 dB SPL calibration tone.
Clapping of 100 dB SPL right outside the box remained under the noise floor measured inside the box >20kHz. Also, speaker play-
backs of 116 dB SPL of 70 to 20 kHz pulses failed to trigger the system and when the system was triggered manually did not lead to
any energy above the noise floor.

QUANTIFICATION AND STATISTICAL ANALYSIS

For statistical analysis of the number of sound emissions for the treatment and the control groups (Figure 1F) we used the Wilcoxon
rank-sum test. To compare our classifier to random result (Figure 2B), we used the binomial probability distribution function (PDF) and
calculate the probability to get the classifier accuracy or higher randomly for each group. To compare the results obtained when using
scattering network26 for feature extraction to the results obtained when using MFCC34 or basic feature29,30 extraction methods (Fig-
ure 1F), we used Wilcoxon sign rank test with Holm-Bonferroni correction. To test the success in distinguishing between drought-
stressed and control plants (Figure 2) we used Fisher’s exact test. To test the differences in the number of emitted sounds per
day (Figure 3A) we used Wilcoxon signed-ranks tests with Holm-Bonferroni adjustment for multiple comparisons.

e4 Cell 186, 1328–1336.e1–e4, March 30, 2023

 ll
Article OPEN ACCESS

Supplemental figures
A

Power / frequency (dB/Hz)

B C

D E
SVM-Scattering Network 1 Raw
Balanced Accuracy (%)

Balanced Accuracy (%)

SVM-Scattering Network 2
SVM-Scattering Network 3 Normalized
SVM-Scattering Network 4
SVM-Scattering Network 5
SVM-Scattering Network 6

(legend on next page)

 ll
OPEN ACCESS Article

Figure S1. Plant sounds recorded in the acoustic box, related to Figure 1
(A) Plant sound spectrograms. Spectrograms of sounds emitted by stressed plants: a drought stressed tomato, a drought stressed tobacco, a cut tomato and a
cut tobacco. The corresponding time signals and spectra of these sounds are shown in Figures 1C and 1D.
(B) Mean spectra of the sounds recorded in the acoustic box. Average signal spectrum of five groups of normalized sounds: Emitted by dry tomatoes, cut to-
matoes, dry tobacco, cut tobacco, and electronic noise. The light blue areas around the line represent the standard error of the mean.
(C) Example of electronic noise. Top—time signal of the noise, normalized to the peak. Bottom—spectrum of the sound shown above. The electronic sound was
recorded in the acoustic box, but only by one microphone, and was thus not considered as ‘‘plant sound’’ for further analysis.
(D) Comparison of different scattering network configurations. The accuracy of sound classification with 6 different configurations for the scattering network,
using SVM as classifier. Each line represents a different feature set, all obtained by scattering network. The scattering networks had different time intervals,
different Q-factors, and a different number of filters, for exact values see Table S1.
(E) The accuracy of sound classification achieved by CNN models. The balanced accuracy of sound classification achieved by binary CNN models (see STAR
Methods), for training and testing on the raw sound-vectors (black) and for training and testing on sound-vectors that were normalized by dividing each value in a
vector by the vector’s maximal absolute value. Each balanced accuracy result is based on leave-one-plant-out cross validation (LOPOCV, see STAR Methods).
 ll
Article OPEN ACCESS

Figure S2. Analysis of plant sounds recorded in the greenhouse, related to Figure 2
(A) Success of CNN models in distinguishing between tomato sounds and greenhouse (GH) noises. Confusion matrices showing the success of trained CNN
classifiers in distinguishing between tomato sounds and GH noises, with different corrections: raw sound-vectors (upper-left), sound vectors that were
normalized by dividing each value in a vector by the vector’s maximal absolute value (upper-right); raw sound vectors that were super-imposed with raw white
noise vectors (WN) from the opposite group (tomato sound + white noise from the GH, GH sound + white noise from the acoustic box, lower-left); normalized
sound vectors that were super-imposed with a normalized white noise vector from the opposite group (lower-right). These confusion matrices are the summation
of the confusion matrices that are produced in the LOPOCV process (see STAR Methods).
(B) Illustration of the construction of super-imposed sounds. First, libraries of white noises (WN) from the acoustic box and from the greenhouse are generated, by
taking ten 2ms segments from the time segment 200-400ms of each sound from the dry tomato library (acoustic box) and from the empty greenhouse noises.
These selected segments appear 100-300ms before the signal that triggered the recording (see STAR Methods in the main text), and they contain white noises
characteristic to the recording environment. We then generate super-imposed sounds, by summing (element-wise) the tomato sound vectors, each with a
randomly selected greenhouse white noise vector, and by summing the greenhouse sounds, each with a randomly selected acoustic box white noise vector. After
the summation we trim the vectors so that all values greater than 1 are set to 1, and all values smaller than
1 are set to
1.
(C) Tomato-classified sounds per hour. We recorded dry and irrigated tomato plants for 1 h in the green house. Then, using a trained Convolutional Neural
Network (CNN) we filtered the greenhouse noises and left only the tomato classified sounds. We here show the number of tomato-classified sounds, recorded
during 1 h in the greenhouse for dry and irrigated tomato plants. The y axis is truncated at 40 for better resolution; plant 22 emitted 95 sounds.
(D) Receiver operating characteristic (ROC) curve presenting the performance of the classification model used to distinguish between dry and irrigated tomato
plants. We plot the true positive rate as function of the false positive rate for all possible decision threshold, where a decision threshold is of the type: above a
certain number ðnc Þ of tomato-classified sounds per hour we classify the plant as dry, otherwise we classify the plant as irrigated.
 ll
OPEN ACCESS Article

Figure S3. Acoustic manifestations of dehydration and other stress factors in multiple plant species, related to Figure 3
(A) The proportion of sounds per day along nine consecutive days of dehydration in tomato plants. The dots represent the average of 23 plants, while the error bars
represent the standard errors. We find significant differences in the proportion of emitted sounds between the first and second day, and between the second
and third day (p values are <0.01 and <0.001 respectively; p values where calculated using Wilcoxon signed-ranks tests, adjusted for 8 comparisons between
pairs of consecutive days using Holm-Bonferroni method). The soil volumetric water content (VWC) of the plants at the beginning of the recording was 0:21 ± 0:03
ðm3 =m3 ; mean ± std).
(B) The proportion of sounds per hour is plotted as a function of the time of day. Each dot represents the average proportion of sounds emitted in the relevant hour
over all 23 plants and nine days, while the error bars represent the standard errors. Dark gray areas represent the hours of complete darkness, light gray areas
show when the greenhouse lighting was on, and the white area represents the approximate hours of natural daylight.

(legend continued on next page)

 ll
Article OPEN ACCESS

(C) A bar plot showing the proportion of emitted sounds as a function of the plant VWC during sound emission. The bars present the averages of the 23 plants,
while the bars represent the standard errors.
(D) Confusion matrix showing the success of the trained CNN classifiers (see STAR Methods) in distinguishing between sounds emitted by a plant experiencing
VWC<0.05 and sounds emitted by a plant experiencing VWC>0.05, in a LOPOCV process (see STAR Methods). The balanced accuracy score is 72.5%.
(E) The success of CNN models in distinguishing between sounds emitted by tomato plants with low and higher VWC values without pre-classification. Confusion
matrices showing the success of CNN classifiers in distinguishing between sounds that are linked to a plant experiencing VWC<0.01, sounds that are linked to a
plant experiencing VWC>0.05 and sounds recorded in an empty greenhouse (GH sounds). The analysis in all four sub-panels is based on the recordings of the
multi-day experiment (Figure 3). The recorded sounds (either emitted by a plant, or greenhouse noises) were tagged according to the VWC of the plant at the time
of the sound emission. These sounds, along with the empty GH sounds library, were fed to 3-class CNN classifiers. Four pre-processing procedures were used,
and a confusion matrix was calculated for each procedure, while applying LOPOCV process (see STAR Methods): all sounds from the GH experiment were
included and the raw sound vectors were used (upper-left); all sounds from the GH experiment were included and the sound vectors were normalized (upper-
right); sounds from the GH plant-recordings were pre-classified by our CNN, and only sounds classified as tomato sounds were used, as raw sound vectors
(lower-left); sounds from the GH plants-recordings were pre-classified by our CNN, and only sounds classified as tomato sounds were used. All the sounds were
normalized (lower-right). (F) Examples of recorded sounds from different plants. Left - Examples of time signals of 6 plants (top to bottom): Triticum aestivum
(Wheat, dry), Zea mays (Corn, dry), Vitis vinifera (Grapevine, Cabernet Sauvignon variety, cut), Mammilaria spinosissima (Pincushion Cactus, cut), Lamium
amplexicaule (Henbit deadnettle, cut), and Solanum lycopersicum (Tomato, infected with TMV). Right - Spectra of the sounds on the left, normalized. Each of
these plants emitted a significant amount of sounds when stressed, suggesting that sound emission is common under stress among non-woody plants. We have
not succeeded recording from woody plants (almond tree, woody parts of grapevine) using our methodology.
 ll
OPEN ACCESS Article

Figure S4. Additional investigation of the mechanism of plant sound emission, related to Figure 4
(A) Sound frequency and trachea cross-section width. Boxplots of the distribution of the plant’s max frequencies vs. its tracheas’ cross section width (estimated
as the average of the ellipse width and height, see examples below) for grapevine, tobacco, and tomato.
(B) Examples of three photos used for the estimation of trachea dimensions. Grapevine (Left) Tobacco (middle), and tomato (right). We performed horizontal cross
section of tomato, tobacco and grapevine stems to measure their trachea cross section area from pictures taken under the microscope, as shown, using
ImageJ.59
(C) Classification of sounds emitted by plants that experienced both dehydration and cutting. The sounds emitted by cut-and-dry plants in the acoustic box were
classified as either ‘‘cut’’ or ‘‘dry’’. We see that while the overall fraction of sounds classified as ‘‘dry’’ in that experiment is around 55%, it varies with time: right
after cutting, sounds classified as ‘‘cut’’ are more common, but from day 2 the majority of sounds are classified as ‘‘dry’’.
(D) Comsol simulations demonstrating sound propagating in all directions from within a stem-like structure. Presented is the 3D far-field sound field.
(E) Horizontal cross section through the sound field from (D) (different colors show different heights along the stem). The simulation shows that sound spreads
equally in all directions in the horizontal plain. The virtual model included two concentric silicon tubes (3.6cm diameter and 4 cm diameter) representing the xylem
vessels within the stem. Because the size was ca. 10 times larger than a real plant, we used a frequency ca. 10 times lower – 7500Hz, thus maintaining the ratio
similar to the real plant. The length of the stem model was 45cm to avoid any effects of length.