Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

S07: Image-based Analysis of Water Content

Change in Tomato Cultivation


Shuto Namba, Junpei Tsuji, Masato Noto
Dept. of Electrical, Electronics and Information Engineering,
Kanagawa University, Yokohama, Japan
s-namba@nt.ee.kanagawa-u.ac.jp, {tsuji, noto}@kanagawa-u.ac.jp

Abstract—Water-stress cultivation is becoming widely used for stress has not been clarified. We must reproduce water-stress
deliberately reducing the amount of irrigation and increasing cultivation techniques carried out by experienced farmers for
the sugar content of fruits. Deliberately reducing the amount of new farmers to inherit them. The ultimate goal of our research
irrigation requires experience and intuition of skilled farmers, so
it is difficult for new farmers to conduct water-stress cultivation. is to develop an automatic irrigation-control system that can
New farmers need techniques to conduct such cultivation easily. be used for water-stress cultivation. Therefore, we must clarify
Under these circumstances, the importance of an irrigation- the relationship between plants and irrigation.
control method for automating irrigation should be discussed. Plants have an internal state called body-water content,
To conduct irrigation control, we must determine the changes in and changes in such content cause water stress. Unless we
plant states. In this study, we clarified how much the body-water
content of plants changes due to irrigation. We also designed a clarify changes in the body-water content of plants, water-
model for representing the body-water content of plants and their stress control will remain difficult. Therefore, it is necessary to
response to irrigation using a hidden Markov model and discuss design a model representing the body-water content of plants
whether it can be used for irrigation control. Such technology for and their response to irrigation. Image-processing techniques
new farmers should be low in cost, so plant reaction to irrigation to determine plant responses to irrigation should be passed
is measured using optical flow on the basis of leaf wilt. Based on
the displacement vectors obtained from optical flow, we estimated down to new farmers. There is no established method for
the body-water content of plants and evaluated whether it is measuring the body-water content of plants from image data,
possible to determine the irrigation amount with our model. so it is necessary to determine the internal state of plants using
The results from experiments conducted on the cultivation of image-processing technology to achieve irrigation control.
tomato plants revealed that the displacement vectors obtained In this study, we designed a model that involves image-
from optical flow under specific conditions reflects the change in
body-water content. processing technology for measuring the body-water content
Index Terms—water-stress cultivation, irrigation, body-water of plants and their response to irrigation. In this paper, we
content of plants discuss whether this model can be used for an irrigation-
control system that we aim to develop.
I. I NTRODUCTION
II. R ELATED WORK
Advanced cultivation techniques to improve the quality of
crops is spreading in Japan. However, as farmers age, the A. Overview
agriculture turnover rates will increase, so there is concern The formalization of techniques used by experienced farm-
that the techniques of skilled farmers will be lost before new ers, such as automatic irrigation-control systems, and repro-
farmers can inherit them. Therefore, research on the spread ducing these techniques through information communication
of advanced cultivation techniques has been conducted. Such technology is known as smart agriculture. Smart agriculture
studies have attempted to reproduce tacit knowledge of skilled results in high agricultural productivity, cost reduction, and
farmers based on environmental and image data [1]–[3]. safer food and working conditions [9] [10]. It consists of
Water-stress cultivation, which is done to deliberately re- systems based on the Internet of Things, sensors, cloud com-
duce the amount of irrigation based on the experience and puting, mobile computing, and big data analysis. To construct
intuition of skilled farmers, is known as an advanced culti- an automatic irrigation-control system, sensing technology to
vation technique. Worldwide, water stress due to drought is determine plant states is indispensable. As sensing technology
recognized as the most dangerous stress to plants [4]. develops, controlling becomes easier with regard to plant
In Japan, however, it is used to produce fruits with high conditions. This section introduces how sensing technology
sugar content and improve market value. To reproduce the is used for agriculture.
tacit knowledge of advanced cultivation techniques in order for
new farmers to easily inherit them, studies on estimating water B. Field Server
stress have been conducted. There are methods for measuring A field server is used for controlling and measuring field
soil-moisture content [5], [6] and body-water content of plants information in open-field cultivation [1]. It consists of sensors,
directly [7] and indirectly [2], [3], [8]. However, how skilled computers, and communication devices as one module. By
farmers carry out irrigation based on the estimated water setting up many field servers, a wireless network using built-in
wireless LAN is formed. With this wireless network, a field- the water potential, and it has been shown that such changes
monitoring system for collecting data measured at each point are effective as an indicator of water stress. In addition,
in real time is constructed. It can be used for managing soil- the linear variable displacement transducer (LVDT) used to
moisture information and the growth process of agricultural measure stem diameter measures the contour of an object, so
crops. Since a field server can be used as a communication it can be used to assess water stress nondestructively, without
infrastructure, development of a new agricultural form is also contact, and continuously. However, as the stem hardens with
expected [11]. However, it is difficult for farmers to maintain age and gradually stops changing, correlation between stem-
field servers due to their complex structure; thus, long-term diameter change and water potential decreases with time. In
use is an issue. long-term cultivation, there is a limit to evaluating water stress
when measuring at the same location. Farmers need to change
C. Plant-disease Detection the location of measuring instruments at regular intervals.
Due to the recent spread of smartphones, they are expected Furthermore, an expensive measuring device, e.g., an LVDT,
to be used for plant-disease detection. Since smartphones is required to measure minute changes in stem diameter. Thus,
can be easily introduced, disease diagnosis is becoming a it is expensive and difficult for ordinary farmers to introduce
technology that anyone can use. As a result of deep learning easily.
using a public dataset of diseased and healthy plant leaves, it
has become clear that a trained model can achieve an accuracy G. Estimation of Water Stress by Using Leaf Wilt
of 99.35% for a test set [12]. However, estimation accuracy is Research using image-processing technology has been con-
low in an environment that differs from that used for training. ducted for estimating water stress. Such technology involves
Disease diagnosis is also used as a function of agricultural using a camera that is easy to introduce, even for new
robots [13]. Therefore, further development of precise sensing farmers, without using expensive sensors [2] [3]. Since image-
technology is expected. processing technology targets the appearance of plants, water
stress is estimated from the leaf wilt.
D. Estimation of Water Stress from Soil-moisture Content Studies on estimating water stress by monitoring the pro-
Water stress is known to occur due to an imbalance between jected plant area by focusing on leaves have been reported [2].
the amount of water absorbed by plants and the amount of When controlling the addition of water stress, it is desirable
water released. The soil-moisture content is closely related to that such stress be detected as soon as possible. From produc-
the amount of water absorbed by plants and measured using a ing images of projected areas set to the horizontal plan 0, 30,
soil-moisture meter. A soil-moisture meter has a tensiometer 60, and 90◦ , it was found that water stress can be detected
[5] for measuring moisture tension, which indicates the ability more quickly in a projected area taken from 90◦ . Introduction
of soil particles to absorb water, and a dielectric soil-moisture into an actual cultivation environment is difficult since the
meter [6] for measuring the volume occupied with water in a image background is a blue sheet.
certain volume of soil. Tensiometers are suitable for measuring Optical flow, which is conducted for motion detection in
the amount of water that plants can use. However, such a image processing, has been reported as a suitable technique in
water-stress-assessment method has a trade-off of accurate actual cultivation environments [3]. There are correlation [15]
measurement but increased cost. Thus, easy introduction to and gradient [16] methods for optical flow detection. With
cultivation environments is difficult. the gradient method, it is assumed that the brightness of the
corresponding point between images does not change because
E. Estimation of Water Stress by using Pressure-chamber of object movement. It is possible to detect even a small
Method change in the motion of an object to be tracked. According
One method for assessing water stress from directly measur- to the study conducted by Kaneda [3], water stress can be
ing the water content in plants is the pressure-chamber method estimated from wilted leaves using optical flow.
[7]. It is a method of applying pressure to leaves collected as
III. P ROPOSED M ODEL
samples and measuring the water from the leaves as the water
potential. The water potential indicates the moisture-holding A. Plant Model for Irrigation-control System
power of a plant and is used as an indicator of water stress. Controlling irrigation based on estimated water stress is
Therefore, the pressure-chamber method can be used to assess indispensable for new farmers to carry out water-stress cul-
water stress with high accuracy. However, this method requires tivation. When considering the passing on of techniques to
collecting leaves, which puts a burden on farmers and increases new farmers, the cost of establishing an irrigation-control
cost. There are also problems with its introduction to actual method should be low. We used image-processing technology
cultivation environments in terms of operation. to measure without leaf wilt using expensive sensors for water-
stress cultivation and optical flow that can be used to detect
F. Estimation of Water Stress by Stem-diameter Change the withering of small leaves.
One method for assessing water stress by indirectly measur- Several studies have proved that water stress on plants can
ing the body-water content of plants is to use stem diameter be estimated from leaf wilt. It is thought that controlling leaf
[8], [14]. Changes in stem diameter strongly correlate with movement due to irrigation may lead to effective management
of water stress on plants. The following three items are displacement vectors obtained from conducting optical flow at
necessary to make irrigation control on plants possible under times t, time t + ∆. The case in which Yt is moving upward
water stress. is considered a state in which plant leaves are recovering.
1) To design a model representing the internal state of Therefore, the RTI is required as follows. A set of whole
plants and their response to irrigation. displacement vectors is defined as Vt , and the RTI is defined
2) To observe plant response to irrigation by using a sensor as follows with regard to Vt using the number of upward-
system. direction vectors of displacement.
3) To determine the appropriate amount of irrigation based #{v ∈ Vt | 0 ≤ arg v < π}
on information from the sensor system. rt = (5)
#Vt
The proposed model based on the above items is described
We can say that it is possible to conduct irrigation control
below.
using rt by indicating that it is effective. This study will enable
B. Model Design a new farmer to inherit a skilled farmer’s experiences and
Regarding item 1), the amount of water content in plants can intuition.
be considered as a factor for determining the state outside the E. Observation Environment
plant. The amount of water content in plant changes over time,
We used a greenhouse kit (Length:100 cm×Width:100
and the current body-water content is determined from the past
cm×Height:200 cm) and light-emitting diodes (LEDs) for
body-water content and amount of irrigation. Our plant model
plant cultivation, as shown in Figs. 1 and 3 (growing is
is a hidden Markov model. Let Xt be the body-water content
possible even in an environment using LEDs [17]).
of plant at time t, Ut the amount of irrigation at time t, and
Yt leaf wilt appearing the plants, Xt and Yt can be formulated
as follows. FP
Xt = f (Xt−1 , Ut−1 , wt−1 ) (1)
Yt = g(Xt , εt ) (2)
The notations wt−1 and εt are stochastic variables that define FP
stochastic factors.
C. Measurement with Sensor System 7RPDWR FP
6FUHHQ &DPHUD
Regarding item 2), we focused on leaves by using agricul-
tural image data, so we considered the relationship between
Yt and the image data It from observing Yt . It is difficult to
directly obtain the information of Yt from It , and only partial FP
information of Yt is obtained.
It = P roj(Yt ) (3)
Fig. 1. Greenhouse kit (top view)
Furthermore, optical flow uses the difference between image
data It−1 and It , the following expression is obtain. Images are taken with a web camera (Logicool C270)
rt = Opt(I0 , I1 , . . . , It ) (4) installed so that the distance to tomato plants are 60 cm.
The images include the entire plant, as shown in Fig. 4. The
Let rt be the reaction of plant by irrigation. Therefore, the images are purple because of the LEDs. We found from the
information of Xt at each time is represented by rt through the actual cultivation that optical flow is affected by noise from
filter by g, P rog, Opt, and Xt can be estimated by calculating the LEDs. Therefore, to alleviating such noise, a black screen
rt . was set as the background, as shown in Fig. 4.
D. Determination of Appropriate Irrigation F. Calculating RTI Using Optical Flow
Regarding item 3), on the basis of Ut to maintain an Figure 2 show flow until obtaining the RTI. The image taken
appropriate Xt , rt has the following problems. at time t was regarding It , and that taken at time t + ∆ was
• It is not known whether rt can be used to detect plant regarding It+∆ . Since we processed delta as ∆ = 0.5, the
response to irrigation. result of conducting optical flow on the It and It+0.5 images
• There is no known appropriate design method of rt . is denoted as Vt,t+0.5 . We then updated Vt,t+0.5 sequentially
We designed rt independently as an index for indirectly and calculated the RTI. We repeated optical flow at 0.5-hour
measuring body-water content under irrigation conditions such intervals from 8:00 to 24:00 each day. Because optical flow is
as the amount and timing of irrigation. calculated based on temporal and spatial changes in in light
For the remainder of this paper, we refer to rt as the and shade between images, motion other than the wilting of
recovery trend index (RTI), which is calculated using the leaves may be captured. Therefore, we decided to remove
LPDJHV RSWLFDOsRZ OLQHFKDUW
57,

‫؞؞؞‬
,W
9WW UWW


,W
9WW

57,
UWW
 WLPH>KRXU@

,W 9WW UWW



‫؞؞؞‬

‫؞؞؞‬
Fig. 2. Calculation flow

FP
/('V
FP

&DPHUD

FP

7RPDWR

Fig. 4. Environment for image capturing

Fig. 3. Greenhouse kit (side view)


To measure change in the Xt , we needed to determine the
necessary irrigation time before conducting the experiment.
the displacement vectors with magnitude 3 or less the Vt Therefore, we included a day without irrigation to determine
beforehand and calculated the RTI. The calculated RTI values the change in leaves when there was no irrigation. In con-
are shown in the line chart. firming the change in Yt during irrigation, it is important to
determine whether Xt decreases or increases. Irrigation was
IV. E XPERIMENT not carried out between 8:00 and 24:00 for the day without
irrigation, which was photographed with the web camera. The
A. Overview
line chart in Fig. 5 shows the displacement vectors, in which
The proposed model for estimating the change in Xt as- leaf wilt moved upward, from the Vt on the day without
sesses whether the RTI reflects such changes due to irrigation. irrigation.
We conducted experiments to assess whether the RTI is It is evident from Fig. 5 that leaf wilt did not move upward
effective as an index to reflect the change in the Xt of tomato from 10:30 to 12:00 and 13:00 to 22:00 in the absence of
plants. irrigation. Therefore, these irrigation conditions clarified the
molecule of rt

57,
time [hour] WLPH>KRXU@

Fig. 5. Day without irrigation Fig. 6. One time irrigation

change in Xt from irrigating at 12:00. The following three


experimental conditions were applied on the basis of the
irrigation at 12:00.
• The condition under which irrigation was carried out

57,
once.
• The condition under which irrigation was carried out
more than once.
• The condition under which the irrigation amount varied.

To determine whether RTI can be used to reflect the change in


Xt due to Ut , we carried out irrigation once. Since irrigation
can be carried out multiple times in water-stress cultivation,
we also carried out irrigation twice. Finally, because it is not
WLPH>KRXU@
practical to provide the same amount of water, we not only a
used a fixed Ut but also varied Ut . We assessed whether RTI
Fig. 7. Multiple irrigation
is valid under each condition.

B. One-time Irrigation D. Divided Irrigation


We clarified whether change in Xt can be confirm from the We compared the RTI values when irrigation was carried
RTI when irrigation was conducted at 12:00. To determine if out with varied Ut . To determine whether the difference in Ut
RTI can be used to reflect the change in Xt due to irrigation, causes change in Xt . We prepared Ut of 200, 300, 400, and
we compared the conditions of irrigation and without irriga- 500 ml, and the results of irrigating in 30-minute intervals
tion. Figure 6 shows the RTI values with a Ut of 200 ml at from 12:00 to 16:00, where a marked change occurred, are
12:00 and when not irrigating. shown in Fig. 8.
The graph indicates that the RTI increased after irrigation, As a result of preparing four Ut , the RTI showed that the
suggesting that it is possible to reflect changes in Xt based leaves recovered quickly and grew larger as Ut increased. This
on displacement vectors. confirms that RTI can be used to measure even changes in Ut .
C. Multiple Irrigation Times V. D ISCUSSION
We then assessed whether the RTI can be used to reflect the We conducted irrigation at 12:00 so that we could compare
change in Xt when increasing the number of irrigation times. the RTI with that without irrigation and confirmed whether
We used a Ut of 200 ml at 11:00 and 14:00 and 200 ml at there was a change (Section IV-B). There was a difference in
12:00 and 18:00 on another day. The results are shown in Fig RTI between the two conditions, suggesting the possibility of
7. reflection the change in Xt with the RTI. Irrigation in a state
Compared with no irrigation, we confirmed that RTI can be in which leaves wilt is a case in which Xt decreases and the
used to reflect the change in Xt after carrying out irrigation change in Xt due to irrigation is easy to confirm.
once. On the other hand, for two irrigation times, 14:00 and We also carried out irrigation from a state in which Xt
18:00, the RTI could not be used to measure this change was thought to be increasing and clarified whether changes in
because it did no change. leaves can be reflected using the RTI (Section IV-C). It became
[3] Y. Kaneda, S. Shibata, and H. Mineno, “Multi-modal sliding window-
based support vector regression for predicting plant water stress,”
Knowledge-Based Systems, vol. 134, pp. 135–148, 2017.
[4] N. E. Williams and A. Carrico, “Examining adaptations to water stress
among farming households in Sri Lanka’s dry zone,” Ambio, vol. 46,
no. 5, pp. 532–542, 2017.
57,

[5] L. Richards and W. Gardner, “Tensiometers for measuring the capillary


tension of soil water,” Journal of the American Society of Agronomy,
vol. 28, no. 5, pp. 352–358, 1936.
[6] G. C. Topp, J. Davis, and A. P. Annan, “Electromagnetic determination
of soil water content: Measurements in coaxial transmission lines,” Water
Resources Research, vol. 16, no. 3, pp. 574–582, 1980.
[7] J. Boyer, “Leaf water potentials measured with a pressure chamber,”
Plant Physiology, vol. 42, no. 1, pp. 133–137, 1967.
[8] M. Gallardo, R. Thompson, L. Valdez, and M. Fernández, “Response of
stem diameter variations to water stress in greenhouse-grown vegetable
crops,” The Journal of Horticultural Science and Biotechnology, vol. 81,
WLPH>KRXU@ no. 3, pp. 483–495, 2006.
[9] F. TongKe, “Smart agriculture based on cloud computing and IoT,”
Journal of Convergence Information Technology, vol. 8, no. 2, 2013.
Fig. 8. Divided irrigation [10] H. Channe, S. Kothari, and D. Kadam, “Multidisciplinary model
for smart agriculture using Internet-of-Things (IoT), sensors, cloud-
computing, mobile-computing & big-data analysis,” Int. J. Computer
clear that it is difficult to reflect such change due to multiple Technology & Applications, vol. 6, no. 3, pp. 374–382, 2015.
[11] M. Hirafuji, “Creating comfortable, amazing, exciting and diverse lives
irrigation times by just calculating RTI. Thus, we considered with CYFARS (CYber FARmerS) and agricultural virtual corporation,”
the reasons the RTI cannot be used to reflect changes in Xt . in Proc. of the Second Asian Conference for Information Technology in
Since rt was designed independently, it is difficult to assess Agriculture, 2000, pp. 424–431.
[12] S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using deep learning for
whether it is optimal for reflecting the change in Xt . However, image-based plant disease detection,” Frontiers in plant science, vol. 7,
it is obvious from the results we obtained thus far that rt can pp. 1–10, 2016.
be used to reflect Xt if done during the first irrigation of at [13] S. Arima, S. Shibusawa, N. Kondo, and J. Yamashita, “Traceability
based on multi-operation robot; information from spraying, harvesting
least one day. and grading operation robot,” in Proceedings of 2003 IEEE/ASME
We determined that it is difficult to reflect the change in Xt International Conference on Advanced Intelligent Mechatronics, vol. 2,
through multiple irrigation times with the RTI, we used four 2003, pp. 1204–1209.
[14] Z. Meng, A. Duan, D. Chen, K. B. Dassanayake, X. Wang, Z. Liu,
Ut at 12:00 and clarified whether each Ut caused a change in H. Liu, and S. Gao, “Suitable indicators using stem diameter variation-
Yt afterwards (Section IV-D). derived indices to monitor the water status of greenhouse tomato plants.”
PLoS ONE, vol. 12, no. 2, pp. 1–15, 2017.
VI. CONCLUSION [15] T. Camus, “Real-time quantized optical flow,” Real-Time Imaging, vol. 3,
no. 2, pp. 71–86, 1997.
We proposed a plant model for developing an irrigation- [16] M. Tao, J. Bai, P. Kohli, and S. Paris, “Simpleflow: A non-iterative,
control system for water-stress cultivation. To automate irri- sublinear optical flow algorithm,” in Computer Graphics Forum, vol. 31,
gation carried out by skilled farmers, we discussed whether no. 2, 2012, pp. 345–353.
[17] G. D. Massa, H.-H. Kim, R. M. Wheeler, and C. A. Mitchell, “Plant
irrigation control can be conducted on the basis of our plant productivity in response to led lighting,” HortScience, vol. 43, no. 7, pp.
model. As a result, the state in which the body-water content 1951–1956, 2008.
of plants is thought to be decreasing can be used to determine
the change in body-water content due to irrigation. The RTI
changed due to the difference in irrigation amount when
irrigation control was conducted, which suggests that a pattern
of irrigation amount can be used. However, irrigation must
be carried out more than once since the RTI cannot be used
to reflect changes in body-water content in a state in which
this content is considered to be increasing. The accuracy of
RTI can be improved using environmental information such as
temperature and humidity. In addition, when irrigation control
is conducted on the basis of estimated water stress, we do
not clarified the relationship between water stress and RTI. A
further study on estimating water stress and on RTI should be
conducted.
R EFERENCES
[1] T. Fukatsu and M. Hirafuji, “Field monitoring using sensor-nodes with
a web server.” JRM, vol. 17, no. 2, pp. 164–172, 2005.
[2] K. Takayama and H. Nishima, “Early detection of water stress in tomato
plants based on projected plant area,” Environmental Control in Biology,
vol. 45, no. 4, pp. 241–249, 2007.

You might also like