Prediction Method of Plant Irrigation Timing

Considering Data Imbalance

Ryosuke Mizuno

, Masaya Goto

and Hiroshi Mineno

1,2

Graduate School of Integrated Science and Technology, Shizuoka University, Japan

Research Institute of Green Science and Technology, Shizuoka University, Japan

Keywords: Imbalanced Data, Internet of Things (IoT), Artificial Intelligence (AI), Agriculture, Water-saving

Cultivation.

Abstract: Predicting the plant irrigation timing is an essential task in the domain of agriculture. A model that can

predict the irrigation timing in tomato cultivation can assist new farmers who do not have sufficient

experience and intuition. In this study, we propose an irrigation timing prediction method based on past

irrigation data, environmental data, and plant water stress using a Random Forest model, which is a general

machine learning method. Our proposed model reproduces irrigation decision making by an expert farmer

for new farmers. Furthermore, we propose a method for resolving imbalances, focusing on the change in the

characteristics of the state of plants due to irrigation. This is because irrigation timing data has a large

imbalance, which is known to be difficult to formulate. Our proposed model clarifies the characteristics of

the irrigation class, and can suppress its misjudgment. We evaluated the proposed method using tomato

cultivation greenhouse data in Shizuoka, Japan. The results show a recall of 92% and f-measure 69% and

hence, the irrigation timing can be predicted with high accuracy. In addition, the results show that the model

works effectively to automatically determine the irrigation timing in greenhouse tomato cultivation.

1 INTRODUCTION

The internet of things (IoT) and artificial intelligence

technology have been advanced and spread, and

computers now support human decision making. In

the domain of agriculture, several studies have been

conducted to address the problems arising from the

aging of the farmer population and the lack of heirs.

These studies can be categorized as: studies to

support the work process of farmers using

technology (Vasconez, Kantor, & Auat Cheein,

2019), and studies to formulate and mechanize the

decision making of farmers (Yukimasa et al., 2017;

Navarro-Hellín et al., 2016).

Studies to support and mechanize the work

performed by farmers use sensors, robots, and IoT

technology to make farming efficient. For example,

a farmer can monitor and control a farm without

even visiting it by checking and controlling the

sensors installed on the farm through the web or a

smartphone (Capraro, Tosetti, Rossomando, Mut, &

Serman, 2018; Joaquín,Gutiérrez, Jua, Francisco,

Aracely, & Miguel, Porta-Gándara, 2015). In

addition, by using autonomously operated tractors

and drones, crops can be harvested without the

farmer’s effort and agricultural chemicals can be

efficiently sprayed with little effort (Vasconez et al.,

2019).

Studies to formulate and mechanize decision

making of farmers reproduce advanced cultivation

techniques based on farmer’s experience and

intuition. For this, the plant status, which is

complex, is analyzed and quantified using various

sensing data, such as temperature, humidity,

scattered light, plant image, evapotranspiration, and

plant water stress. In particular, a few studies z

(Yukimasa et al., 2017; Liu et al., 2017; Peng et al.,

2019) have formulated the decision-making with

small and frequent irrigation which is known as

water stress cultivation. This is convenient to

automatically cultivate high-quality fruits and crops.

Peng et al. (2019) proposed a crop water demand

prediction system by using the back propagation

(BP) neural network. The BP neural network was

trained using various environmental data such as the

solar radiation, soil moisture, soil electrical

conductivity, and temperature. The water demand

was evapotranspiration calculated by the Penman-

Mizuno, R., Goto, M. and Mineno, H.

Prediction Method of Plant Irrigation Timing Considering Data Imbalance.

DOI: 10.5220/0009339801770184

In Proceedings of the 9th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2020), pages 177-184

ISBN: 978-989-758-418-3

177

FAO formula (Liu et al., 2017). In order to realize

automatic cultivation, it is necessary to determine

the predicted water demand threshold based on

appropriate irrigating timing. Determining the

threshold value is difficult for new farmers because

this requires experience and intuition cultivated over

a long period of time. Yukimasa et al. (2017)

proposed a model for predicting future plant water

stress by using the Sliding Window-based Support

Vector Regression (SW-SVR). The method was

evaluated using environmental data inside the

greenhouse and image data being generated from the

movement of plant leaves. This study made it easy

to understand future water stress using a simple and

economical sensor. In addition, the model can

understand the water stress from the data of the

cultivation environment. Therefore, not only expert

farmers, but also, new farmers who do not possess

sufficient experience of cultivation can understand

the water stress with the prediction of future water

stress by this method. However, in order to realize

automatic cultivation, it is necessary to determine

the predicted water stress threshold based on

appropriate irrigating timing likewise Peng et al.

(2019).

We propose an irrigation timing prediction

method based on past irrigation data, environmental

data, and plant state data by using machine learning.

Our proposed method reproduces irrigation decision

making by experts and helps new farmers.

Furthermore, we propose a method that resolves

imbalances by focusing on the change characteristics

of the state of plants from irrigation. This is because

the small and frequent irrigation data such as that of

the cultivation of tomato and strawberry has a large

imbalance that is known to be difficult to formulate.

The small and frequent irrigation is conducted

approximately 50 times (total time is approximately

50 minutes) during the day, and non-irrigation

accounts for the major part of the day.

This paper brings two key contributions to the

field of agriculture research: (1) New farmers can

achieve automatic cultivation of fruits of high

quality. This is because the proposed model uses IoT

devices in the greenhouse. (2) Our proposed model

leads to the technology development of modeling of

small and frequent irrigation with data imbalance in

the domain of agriculture.

The rest of the paper is organized as follows:

Section 2 presents a discussion of related techniques

of resolving imbalanced data. Section 3 describes

the proposed method. Section 4 presents the results

from the evaluation of the proposed method using

actual agricultural data. Finally, we present the

conclusions and future work in Section 5.

2 TECHNIQUES TO RESOLVE

IMBALANCED DATA

To solve the imbalance of datasets is an important

task in predicting irrigation timing using machine

learning. There are a few methods to solve data

imbalance such as classifier level methods and data

level methods.

Classifier level methods are cost-sensitive

learning methods that vary the error transmitted to

each class. Cost-sensitive learning methods assign

weights to the samples to match a specific data

distribution. Weighting by inverse class frequency

(Chen, Change, & Xiaoou, 2016; Yu-Xiong et al.,

2017) has often been adopted. To rephrase, the

minority data which is difficult to classify, weights

the penalty. In our evaluation, we adopt cost-

sensitive learning, in which the reciprocal of the

ratio of minority data to the number of data points is

multiplied as a penalty for errors in the minority data.

Data level methods are data sampling techniques.

Two types of sampling techniques are shown in

Figure 1: oversampling methods that increase the

number of minority data and undersampling

methods that reduce the number of majority data.

Oversampling methods add or reuse new data to

increase the minority data. Random sampling

repeatedly samples from the minority data. Synthetic

Minority Over-sampling Technique (SMOTE)

(Chawla et al., 2002) and Adaptive Synthetic

Sampling (ADASYN) (Haibo et al., 2008) share the

concept of generating new data on a line connecting

minority data. SMOTE adds a random number

multiplied with the sample on the line connecting

the selected minority samples. In ADASYN, the

value to be multiplied is determined according to the

number of majority data contained in the K

neighbors of the selected minority sample. Therefore,

ADASYN reduces the frequency with which

minority data is generated near the majority data.

However in irrigation timing data, the minority data

is similar to the majority data, and hence, we

consider the data generated near the majority data.

Undersampling methods eliminate a few samples

from the majority data. Random sampling randomly

Figure 1: Addressing imbalanced data with resampling.

Oversampling

Original

(Imbalanced)

Balanced

Undersampling

Original

(Imbalanced)

Balanced

SMARTGREENS 2020 - 9th International Conference on Smart Cities and Green ICT Systems

178

Algorithm 1: Undersampling for eliminating data based on near irrigation timing.

Input:

Imbalanced dataset: 











, 





, …,







,







ℎ 



∈



,



∈

Distance parameter: n, m

Output:

Balanced dataset: 











, 





, …,







,







ℎ 



∈



,



∈

Definition of Undersampling for eliminating data based on near irrigation timing:

For t=1 to N

if 



is irrigation

Eliminate non-irrigation data for period



−





−1



Eliminate non-irrigation data for period



+1





+







← Remaining data not eliminated

determines the samples to be removed from the

majority data. NearMiss algorithm (Jianping &

Inderjeet, 2003; Yen & Lee, 2009) uses the K-

nearest neighbor method to remove clearly

identifiable majority data. This eliminates data that

is difficult to distinguish from between the minority

class and the majority class. Therefore, it is possible

to separate the details of the decision boundary by

using data after applying the NearMiss algorithm.

However, it is difficult to determine the parameters

of the K-neighbor method using NearMiss. This is

because the irrigation timing data has a characteristic

minority data and majority data that are similar. We

consider that undersampling is suitable for resolving

irrigation data imbalance. Therefore, we propose an

undersampling method considering the

characteristics of the irrigation data of plants. This is

because the farmer does not irrigate depending on

the moisture state of the plant even at the same

temperature. Irrigation is performed depending on

the amount of solar radiation and season even in

different plant water states. Therefore, we consider it

better to reduce the data so that the irrigation class

and the non-irrigation class are clear rather than

increasing the data using oversampling or by

considering the cost.

3 MODEL DESCRIPTION

We propose a method for resolving imbalances

suitable for irrigation timing to build the model. We

aim to build the model to predict the irrigation

timing by farmers using the environmental data of

the greenhouse and hereby, reproduce the irrigation

timing automatically. Furthermore, we aim to

automatically cultivate by controlling IoT devices,

which are able to control the irrigation timing in the

greenhouse based on the proposed method. The

process is composed of two main elements to build

the irrigation timing model. First, we address the

imbalance of irrigation timing data by using

undersampling for eliminating data based on near

irrigation timing (ENIT). Next, we build the model

to predict the irrigation timing using the balanced

data after solving imbalance. Section 3.1 presents

the algorithm of ENIT and Section 3.2 presents the

irrigation timing prediction method using machine

learning.

3.1 Addressing Irrigation Data

Imbalance

We address the imbalance in the irrigation timing

data by using undersampling for eliminating data

based on near irrigation timing (ENIT) to eliminate

the data of majority class (non-irrigation data) near

the time of the data of minority class (irrigation data)

(Algorithm 1). It may be noted that the irrigation

timing data is imbalanced because the frequency of

irrigation time is approximately 50 (total time is

approximately 50 minutes) during the day, and non-

irrigation accounts for the major part of the day. One

of the decisions on irrigation timing is to use the

value of solar radiation accumulated from the

previous irrigation (Takayoshi et al., 2018).

Similarly, the value of evapotranspiration is used for

the decision on irrigation timing (Pawlowski et al.,

2017; Peng et al., 2019). Irrigation is performed

when the accumulated value exceeds the threshold.

Therefore, we have considered that the plant state at

Prediction Method of Plant Irrigation Timing Considering Data Imbalance

179

the time of irrigation and the plant state in the past

direction for a certain period of time from the start

of irrigation are similar. The accumulated

environmental data, for example temperature and

evapotranspiration from the last irrigation, are also

assumed to be characteristically similar. In addition,

we have considered that the state of the plants near

irrigation are similar because irrigated plants need

time to absorb moisture from the soil through the

roots and through evapotranspiration from the leaves

to allow water to enter the body. For this reason, we

have eliminated the data of the majority class (non-

irrigation class) which is near the data of minority

class (irrigation timing class) by using ENIT. ENIT

is an undersampling method that removes the non-

irrigation data that is nearer in the time series based

on irrigation timing, and the algorithm is shown in

Algorithm 1. In the ENIT algorithm, when the

duration of irrigation timing is t, the non-irrigation

data represented by (t-n), …, (t-1) is eliminated for

the parameter n in the past direction and the non-

irrigation data represented by (t+1), … , (t+m) is

eliminated for the parameter m in the future

direction. As a result, the non-irrigation class data

that has similar characteristics to the irrigation class

is eliminated by selecting the data from the non-

irrigation class of the majority based on the

irrigation class of the minority. In addition, the

characteristics of the irrigation class are clarified and

misjudgment is suppressed.

3.2 Predicting the Irrigation Timing

We have used Random Forest (RF) to predict the

irrigation timing when irrigation is necessary. The

RF is one of the general machine learning methods

and is an ensemble learning method combining

multiple decision trees. An ensemble learning

method is a modeling method that consists of a

combination of prediction of multiple classifiers

rather than the prediction of a single classifier. By

applying ensemble learning, the predictive value is

diverse and can be predicted robustly for unknown

data. Therefore, we have adopted ensemble learning

in order to make a robust prediction model.

4 EXPERIMENTAL PROCEDURE

In this section, we describe the dataset of the

experiment for evaluation, the experimental

parameters, and the results.

(a) Overview of greenhouse and

cultivation line.

(b) Layout of measurement sensors.

Figure 2: Dataset collection environment.

4.1 Data Collection and Preprocessing

A system was developed to collect agricultural data

in a greenhouse of tomato (Solanum lycopersicum L.

cultivar Frutica) at Fukuroi, Japan. We installed

environmental sensor nodes, laser displacement

sensors (HL-T1010A Panasonic Corporation), and a

datalogger (midi LOGGER GL840 Graphtech

Corporation) on the greenhouse. An overview of the

greenhouse and cultivation line and the layout of

measurement sensors are shown in Figure 2 (a) and

(b), respectively. Environment data such as

temperature, relative humidity, solar radiation, and

vapor pressure deficit (VPD) was collected along

with the data of the plant stem-diameter and

irrigation timing. Drip irrigation was used in which a

certain amount of water was released in one

irrigation. In addition, time series features related to

water stress and tomato irrigation were calculated to

create datasets. Stem-diameter represented the plant

growth and water stress. However, Kazumasa et al.

(2019) showed that the diameter could not be used

Block4

Block5

Block2

Block3

Block6

Block8

Block7

Block9

Block1

Greenhouse A Greenhouse B Greenhouse C

Target

plant

Cultivation line

Laser displacement sensor

Target plant

Rockwool cube

Steel frame

Environmental sensor node

Ground

Cultivation bed

Guide string

SMARTGREENS 2020 - 9th International Conference on Smart Cities and Green ICT Systems

180

as a water stress index directly because it changed

with the growth of the plant and diurnal variation.

Therefore, we have defined the difference in stem

diameter calculated using the most recent irrigation

(DSR) as a water stress index in accordance with the

work of Kazumasa et al. (2019). The DSR is a value

calculated by subtracting the current stem diameter

from the maximum stem diameter. The recent

irrigation in current time t is calculated as follows:













,



,…,





−



. (1)

Where t is the current time and n is the time

elapsed since the most recent irrigation. Additionally,

we have defined the time series features such as the

elapsed time since sunrise, elapsed time since the

previous irrigation, and accumulated environmental

data from the previous irrigation. This is because

Kazumasa et al. (Kazumasa et al., 2018) showed that

the machine learning method without recursion can

be improved by considering the time series features.

Table 1 shows the features of the dataset that were

finally calculated. All data were collected at a

frequency of once every minute and during the

periods A (from March 28 to October 22, 2018 in

greenhouse A), B (from October 23, 2018 to January

16, 2019 in greenhouse B), and C (from April 25 to

June 20, 2019 in greenhouse C). During periods A,

B, and C, the irrigation data was collected for one,

four and four blocks, respectively. Therefore, we

have collected irrigation data for nine blocks, or to

paraphrase, datasets were created for nine different

scenarios.

4.2 Experimental Condition

We have evaluated the performance of the proposed

method using actual agricultural data. In the

evaluation, the prediction accuracy of the irrigation

timing was compared by using environmental data

related to the irrigation of tomato as shown in Table

1. The recall and f-measure were used as error

indicators when the threshold for classification

judgment is 0.5 (50%). Recall shows the rate at

which the irrigation timing predicted by the model

matches the irrigation timing by the farmer. F-

measure shows the accuracy and completeness of the

irrigation timing predicted by the model. The

conditions are detailed in Table 2. There are

seventeen different conditions: no addressing of

imbalance, cost-sensitive learning, three

oversampling methods, two undersampling methods,

with and without ENIT before applying

undersampling method, and changing the

undersampling rate. The important parameters of the

RF method were tuned by using grid search:

Table 1: Features of dataset.

Type Feature

Environmental data

Temperature

Relative humidity

Solar radiation

Vapor pressure deficit

Plant water stress

Stem-diameter

DSR

Time-series feature

Elapsed time since sunrise

Elapsed time since last irrigation

Accumulated Environmental

data

Table 2: Evaluation condition.

Name Address imbalanced

Base No

Cost Inverse Class Frequency

OverRandom RandomSampling

SMOTE SMOTE

ADASYN ADASYN

UnderRandom02

RandomSampling

UnderRandom04

UnderRandom06

UnderRandom08

UnderRandom10

ENIT_UnderRandom10

ENIT &

RandomSampling

NearMiss02

NearMiss-1

NearMiss04

NearMiss06

NearMiss08

NearMiss10

ENIT_NearMiss10

ENIT & NearMiss-1

*: "02" means that the number of minority data is 2 when the

number of majority data 10 and "10" means balanced between

majority data and minority data.

Prediction Method of Plant Irrigation Timing Considering Data Imbalance

181

Table 3: The number of training and validation data

points.

Name

The number of data

points (Non-

irrigation/irrigation)

Base, Cost 231,250 / 11,047

OverRandom,SMOTE,ADASYN 231,250 / 231,250

UnderRandom02, NearMiss02 55,235 / 11,047

UnderRandom04, NearMiss04 27,617 / 11,047

UnderRandom06, NearMiss06 18,411 / 11,047

UnderRandom08, NearMiss08 13,808 / 11,047

UnderRandom10, NearMiss10,

ENIT_UnderRandom10,

ENIT_UnderRandom10

11,047 / 11,047

n_estimators (10, 20, 30) and max_depth (15, 20,

40). In addition, we set the ENIT hyperparameter n

and m to 2.

Evaluation data such as training, validation, and

test data were divided as per the following procedure.

First, the data set was divided into periods A to C.

Next, a day was calculated that included 80% of the

total number of irrigations in each period. The data

of the period after that date was set as the test data.

For training and validation data, the data excluding

the test data was divided into 5 parts, and 5-fold

cross validation was applied. In addition, from the

test and validation data, the majority data was

deleted to random to create equilibrium data to

obtain the correct accuracy. The number of test data

points were 4,896 (of which 2,448 were irrigation

data). The number of training and validation data

points before 5-fold cross validation were shown in

Table 3.

4.3 Results and Discussion

Figure 3 shows the errors of each comparison for the

testing data. The combination of the proposed

method and NearMiss has the highest score for both

recall and f-measure: 0.92 recall and 0.69 f-measure.

Only NearMiss has a score of 0.91 recall and 0.69 f-

measure. In addition, the combination of the

proposed method and random undersampling scores

better than only random undersampling. This is

because, the combination of the proposed method

and random sampling and only random sampling

have a score of 0.78 recall and 0.69 f-measure and

0.71 recall and 0.66 f-measure, respectively. These

Figure 3: The results of each approach for imbalanced

data.

Figure 4: The results of undersampling with changing rate.

results show that the proposed method works

effectively. In addition, the results show that

undersampling is superior to cost-sensitive learning

and oversampling. Random oversampling is

considered to be inaccurate because irrigation data

that provide useful features are not selected. In

SMOTE and ADASYN, the data is generated on a

line connecting minority data. Therefore, minority

data may be generated in the majority area. However,

undersampling by NearMiss, which has the highest

accuracy, does not generate data near the majority

data. NearMiss has the characteristic that the

majority data near the decision boundary is

unchanged when ideal processing is performed.

Therefore, the model using NearMiss learns detailed

decision boundaries and the accuracy is improved.

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

Recall F-measure

Recall / F-meaure

Base Cost OverRandom

SMOTE ADASYN UnderRandom10

ENIT_UnderRandom10 NearMiss10 ENIT_NearMiss10

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

Recall F-measure

Recall / F-meaure

UnderRandom02 UnderRandom04 UnderRandom06

UnderRandom08 UnderRandom10 ENIT_UnderRandom10

NearMiss02 NearMiss04 NearMiss06

NearMiss08 NearMiss10 ENIT_NearMiss10

SMARTGREENS 2020 - 9th International Conference on Smart Cities and Green ICT Systems

182

Figure 4 shows the result of changing the

sampling rate of undersampling. Both random

sampling and NearMiss increase in accuracy as the

sampling rate increases. To paraphrase, the accuracy

is higher for the cases having data closer to the

balanced data. In particular,

ENIT_UnderNearMiss10 has a recall of 0.92 and

can predict irrigation timing with high accuracy.

5 CONCLUSIONS

We proposed a novel method for resolving

imbalances suitable for irrigation timing and its

prediction. We addressed the imbalance of irrigation

timing data by using undersampling for eliminating

data based on near irrigation timing (ENIT), to

eliminate the non-irrigation data near the time of

irrigation. The performance of the proposed method

was evaluated using actual agricultural data. In the

evaluation, the prediction accuracy of irrigation

timing was compared by using environmental data

related to the irrigation of tomato. In the results, The

accuracy was improved by the two methods that

applied the proposed method. We showed that the

prediction accuracy of small frequent irrigation can

be improved by applying the method for eliminating

imbalances that takes into account the characteristics

of irrigation timing data. This result shows that it is

necessary to eliminate the imbalance in the

prediction of irrigation timing. Furthermore, the

result shows that it is effective to consider irrigation

characteristics to eliminate imbalance. The aim in

future is to automatically cultivate various crops by

controlling through IoT devices, which are able to

control the irrigation timing in greenhouses based on

the proposed method. IoT technology has already

been introduced in the agricultural domain.

In future, we will evaluate the general purpose of

the proposed method under various conditions with

different greenhouses, cultivation methods, and

water supply. In addition, the prediction model will

be examined. Specifically, the application of Long-

Short Term Memory (LSTM) (Sepp & Jurgen, 1997),

which is one of the most powerful deep learning

methods, will be considered. LSTM can be

considered for irrigation timing because it can

consider long-term time series. In addition, we will

consider Dynamic Time Warping (DTW) (Bemdt &

Clifford, 1994) to error indicator. Recall and F-

measure are evaluated for one point in time without

considering time series. Thus, a model that is off by

only one point in time and a model that cannot be

predicted at all are both incorrect. Therefore, we

evaluate the similarity between two time-series

sequences using DTW.

ACKNOWLEDGEMENTS

We greatly appreciate Mr. Makoto Miyachi (Happy

Quality Co., Ltd., Japan) and Mr. Daigo Tamai (Sun

Farm Nakayama Co., Inc., Japan) for providing an

environment for data collection.

REFERENCES

Bemdt, D. J., & Clifford, J. (1994). Using Dynamic Time

Warping toFindPatterns in Time Series. AAAI, 359–

370.

Capraro, F., Tosetti, S., Rossomando, F., Mut, V., &

Serman, F. V. (2018). Web-based system for the

remote monitoring and management of precision

irrigation: A case study in an arid region of Argentina.

Sensors (Switzerland), 18(11). https://doi.org/

10.3390/s18113847

Chawla, N. V, Bowyer, K. W., Hall, L. O., & Kegelmeyer,

W. P. (2002). SMOTE: Synthetic Minority Over-

sampling Technique. In Journal of Artificial

Intelligence Research (Vol. 16).

Chen, H., Chen, Change, H., & Xiaoou, T. (2016).

Learning Deep Representation for Imbalanced

Classification. CVPR.

Haibo, H., Yang, B., Edwardo, A, G., & Shutao, L.

(2008). ADASYN: Adaptive Synthetic Sampling

Approach for Imbalanced Learning. 1322–1328.

https://doi.org/10.1109/IJCNN.2008.4633969

Jianping, Z., & Inderjeet, M. (2003). kNN Approach to

Unbalanced Data Distributions: A Case Study

involving Information Extraction. Int’l. Conf.

Machine Learning1(ICML).

Joaquín,Gutiérrez, J., Jua, Francisco, V.-M., Aracely, L.-

G., & Miguel, Porta-Gándara, Á. (2015).

Smartphone Irrigation Sensor. Sensors, 15(9),

5122–5127.

https://doi.org/10.1109/JSEN.2015.2435516

Kazumasa, W., Ryosuke, M., Gota, N., & Hiroshi, M.

(2019). Multimodal neural network with clustering-

based drop for estimating plant water stress.

Computers and Electronics in Agriculture, 105118.

https://doi.org/10.1016/J.COMPAG.2019.105118

Kazumasa, W., Shun, S., Hiroshi, M., Takeshi, S., Daichi,

S., Yoshikazu, K., & Katsumi, S. (2018). Time

series Feature Injection Method for Estimating

Plant Evapotranspiration using Neural Networks.

Pattern Recognition and Media Understanding

(PRMU).

Liu, X., Xu, C., Zhong, X., Li, Y., Yuan, X., & Cao, J.

(2017). Comparison of 16 models for reference crop

evapotranspiration against weighing lysimeter

measurement. Agricultural Water Management,

Prediction Method of Plant Irrigation Timing Considering Data Imbalance

183

184, 145–155. https://doi.org/10.1016/J.AGWAT.

2017.01.017

Navarro-Hellín, H., Martínez-del-Rincon, J., Domingo-

Miguel, R., Soto-Valles, F., & Torres-Sánchez, R.

(2016). A decision support system for managing

irrigation in agriculture. Computers and Electronics

in Agriculture, 124, 121–131. https://doi.org/

10.1016/j.compag.2016.04.003

Pawlowski, A., Sánchez-Molina, J. A., Guzmán, J. L.,

Rodríguez, F., & Dormido, S. (2017). Evaluation of

event-based irrigation system control scheme for

tomato crops in greenhouses. Agricultural Water

Management, 183, 16–25. https://doi.org/10.1016/

J.AGWAT.2016.08.008

Peng, Y., Xiao, Y., Fu, Z., Dong, Y., Zheng, Y., Yan, H.,

& Li, X. (2019). Precision irrigation perspectives on

the sustainable water-saving of field crop

production in China: Water demand prediction and

irrigation scheme optimization. Journal of Cleaner

Production, 230, 365–377. https://doi.org/10.1016/

J.JCLEPRO.2019.04.347

Sepp, H., & Jurgen, S. (1997). Long Short-Term Momory.

Neural Computation, 9(8), 1735–1780.

Takayoshi, Y., Takafumi, K., Makoto, S., & Hiroki, K.

(2018). Effects of ingredient and application

method of controlled-release fertilizers on blossom-

end rot and fruit yield of tomato in the high-density

cultivation with low node-order pinching using

solar radiation powered pulsating drip irrigation

system. Bulletin of the NARO, Agricultural

Research for Western Region, 53–64.

https://doi.org/10.24514/00000237

Vasconez, J. P., Kantor, G. A., & Auat Cheein, F. A.

(2019). Human–robot interaction in agriculture: A

survey and current challenges. Biosystems

Engineering, 179, 35–48. https://doi.org/10.1016/

J.BIOSYSTEMSENG.2018.12.005

Yen, S.-J., & Lee, Y.-S. (2009). Cluster-based under-

sampling approaches for imbalanced data

distributions. Expert Systems with Applications,

36(3), 5718–5727. https://doi.org/10.1016/

J.ESWA.2008.06.108

Yu-Xiong, W., Deva, R., & Martial, H. (2017). Learning

to Model the Tail. NIPS, 7032–7042.

https://doi.org/10.3390/atmos7060083

Yukimasa, K., Shun, S., & Hiroshi, M. (2017). Multi-

modal sliding window-based support vector

regression for predicting plant water stress.

Knowledge-Based Systems, 134, 135–148.

https://doi.org/10.1016/j.knosys.2017.07.028

SMARTGREENS 2020 - 9th International Conference on Smart Cities and Green ICT Systems

184