Computers and Electronics in Agriculture 88 (2012) 1–12
Contents lists available at SciVerse ScienceDirect
Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag
Development of an autonomous early warning system for Bactrocera dorsalis (Hendel) outbreaks in remote fruit orchards Min-Sheng Liao a,1, Cheng-Long Chuang a,b,1, Tzu-Shiang Lin a, Chia-Pang Chen a, Xiang-Yao Zheng a, Po-Tang Chen a, Kuo-Chi Liao a, Joe-Air Jiang a,⇑ a b
Department of Bio-Industrial Mechatronics Engineering, National Taiwan University, Taipei 10617, Taiwan Intel-NTU Connected Context Computing Center, National Taiwan University, Taipei 10617, Taiwan
a r t i c l e
i n f o
Article history: Received 9 February 2012 Received in revised form 24 April 2012 Accepted 28 June 2012
Keywords: Agricultural management Oriental fruit fly Pest monitoring Early warning system Wireless sensor networks
a b s t r a c t Developing an autonomous early warning system for detecting pest resurgence is an essential task to reduce the probabilities of massive Oriental fruit fly (Bactrocera dorsalis (Hendel)) outbreaks. By preventing pest outbreaks, farmers would be able to reduce their dependence on chemical pesticides. Chemical pesticide abuse often brings harmful consequences to human health and natural environments. Since an agroecological system can change at a fast rate due to the soil degradation and the environmental factors changes, the rise of pest density cannot be immediately detected by traditional methodologies. In this study, an autonomous early warning system, built upon the basis of wireless sensor networks and GSM networks, is presented to effectively capture long-term and up-to-the-minute natural environmental fluctuations in fruit farms. In addition, two machine learning techniques, self-organizing maps and support vector machines, are incorporated to perform adaptive learning and automatically issue a warning message to farmers and government officials via GSM networks when the population density of B. dorsalis significantly rises. The proposed system also provides sensor fault warning messages to system administrators when one or more faulty sensors give abnormal readings to the system. Then, farmers and government officials would be able to take precautionary actions in time before major pest outbreaks cause an extensive crop loss, as well as to schedule maintenance tasks to repair faulted devices. The experimental results indicate that the proposed early warning system is able to detect the incidents of possible pest outbreaks in a variety of seasonal conditions with sensitivity, specificity, accuracy, and precision around 98%, 100%, 100%, and 100%, respectively, as well as to transmit the early warning messages to farmers and government officials via Short Message Service using the GSM network. The proposed early warning system can be easily adopted in different fruit farms without extra efforts from farmers and government officials since it is built based on machine learning techniques, and the warning messages are delivered to their mobile phones as text messages. The proposed early warning system also shows great potential to assist farmers to update their pest control operations in the fruit farms, and help government officials to improve farming systems. Ó 2012 Elsevier B.V. All rights reserved.
1. Introduction Recently, the impacts of global warming rapidly change the Earth’s weather system. Climate change and extreme weather events can induce multiple problems in food production, plant diseases, and pest population dynamics (Chen and McCarl, 2001; Fischer et al., 2002; Rosenzweig et al., 2001). Without the proper understanding of the interactions between climate and pests, food producers may be affected in a way that is more disruptive than ⇑ Corresponding author. Address: Department of Bio-Industrial Mechatronics Engineering, National Taiwan University, No. 1, Sec. 4, Roosevelt Rd., Taipei 10617, Taiwan. Tel.: +886 2 3366 5341; fax: +886 2 2362 7620. E-mail address:
[email protected] (J.-A. Jiang). 1 These authors contributed equally. 0168-1699/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.compag.2012.06.008
beneficial because the cost of damage by the pest is more than the cost of control. Since climate is one of the major drivers of pest population dynamics, an innovative and sustainable pest management strategy is required to manage the pests and pathogens with altered status, distribution, and population (Sutherst et al., 2011). In order to enhance the effort to control the pests and pathogens, it is crucial to understand the behavior of complex agroecosystems. Modeling an agro-ecosystem requires analysis of interactive sensing data with high temporal and spatial resolutions. WSNs provide an excellent solution to build a remote agroecological monitoring system that can collect simultaneous multi-sensor data (e.g. local climate parameters and pest population) for modeling the agro-ecosystems with both flexibility and scalability (Bezdek et al., 2011).
2
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
Wireless sensor networks (WSNs) usually comprises a number of wireless sensor nodes distributed over a large area to collect and report sensor measurements, and one or more sink nodes that serve as gateways to organize the sensor nodes into a network, as well as to collect readings that are measured by the sensor nodes. The sensor node is a compact, low-power, and low-cost device that integrates a wireless radio transceiver, a microcontroller, a variety of sensors, and a battery on a small circuit board. Recently, WSNs have been applied to many practical applications to provide high-resolution real-time sensing information about the condition of the physical worlds (Dissanayake et al., 2001; Huang et al., 2006; Subramaniam et al., 2006; Zhang et al., 2010). Despite the rapid growth of WSNs in the last few years, information acquired by WSNs might still be unreliable. This fact is mainly owing to faulty readings of sensor nodes caused by several factors, including the limited power supply, the limited storage or processing capabilities, the intervention of harsh environmental variation, etc. Thus, a wrong decision is usually made whenever the sensing signal is contaminated with outliers. In this study, by combining WSN and Global System for Mobile Communication (GSM) technology, we implemented a remote agroecological monitoring system to assess the climate and pest status of a crop of interest through all seasons with an autonomous event detection algorithm. The primary contributions of this study are: A reliable agroecological monitoring system that utilizes different technologies from mechatronics (multi-disciplines including mechanics, electronics, telecommunication, and information technologies) is presented. The proposed system is able to collect meteorological data and to assess the population density of insect pests in the orchard of interest. A web-based user interface is designed for the system administrator and government agents to access and analyze historical sensing data online via the Internet. The proposed autonomous event detection algorithm is a combination of the self-organizing maps (SOMs) and support vector machine (SVM). Thus, the proposed algorithm can be easily employed in any orchard since it can determine suitable boundaries between different events specifically for the farm of interest. The autonomous event detection algorithm is able to disambiguate agroecological events from sensor node faultiness (e.g. abnormal sensor readings caused by faulted sensor) with high accuracy and low false-alarm rate. When an event is detected, the proposed system will send an alert message to the system administrator if the event is an unusual one. If a pest outbreak occurs, the government agents can activate integrated pest management program as soon as possible in order to reduce economic loss to agriculture. The rest of this paper is organized as follows. Section 2 briefly introduces the architecture of the remote agroecological monitoring system. Section 3 presents the implementation of the autonomous event detection algorithm. Section 4 describes the simulation and seasonal experimental results, and finally, Section 5 devotes to the conclusions.
2. Remote agroecological monitoring system 2.1. Monitoring subject and state-of-art In this study, we implemented a remote agroecological monitoring system that integrates WSN and GSM technologies to
simultaneously record the environmental conditions and pest population. The target pest is Oriental fruit fly (Bactrocera dorsalis (Hendel)). It is one of the most economically essential phytophagous insects in the Asia–Pacific region (Armstrong et al., 2004; Chen et al., 2006; Malacrida et al., 2007; Smith, 1989; Vargas et al., 1984), especially prevalent in Malaya, Thailand, Pakistan, and Taiwan. Hawaii, Mariana Islands, and some parts of continental United States, such as California, also reported its existence (Drew and Raghu, 2002; Weems et al., 2008). Since a female Oriental fruit fly can lay about 1500 eggs during its lifetime, such dispersive capability poses great threats to the agricultural systems in Taiwan (Lin et al., 2005; Metcalf and Metcalf, 1992; Jiang et al., 2008). The fly makes fruits rotten and dropped, so the quality and quantity of fruit production are seriously declined, which can cause severe economic losses. Without adequate control, when the infestation rate reaches 10–30%, the economic losses caused by the Oriental fruit fly can reach 1.4–1.9 billion US dollars per year solely in Taiwan (Hung et al., 2008). The government of Taiwan recognizes the severity of the damage caused by Oriental fruit fly, and annually spends about 2.7 million US dollars to develop a comprehensive control program against it. Starting in 1994 till 2008, more than 61 monitoring stations, each of which includes at least nine spots equipped with fly traps, and thus totally 613 monitoring traps have been set up in Taiwan (Cheng et al., 2002; Agricultural Research Institute, 2010a,b). The number of the trapped flies is counted and recorded every 10 days to analyze the population dynamics of the Oriental fruit fly. However, such surveillance approach relies on manual measurement that can often be incomplete and inaccurate with poor temporal resolution. Furthermore, past studies have demonstrated that the population dynamics of Oriental fruit fly is influenced by the factors like temperature, solar illumination, rainfall volume, and different kinds of crops (Agricultural Research Institute, 2010a,b; Valadão et al., 2010). In order to have an increased knowledge regarding the relation between these factors and the population dynamics of Oriental fruit fly, this study develops a remote agroecological monitoring system. The system is designed to provide simultaneous data (e.g. meteorological data, population dynamics of pest, etc.) with high temporospatial resolution on aspects of the fruit farm. With the information acquired from the system, it is then possible to unravel the underlying mechanisms in the population dynamics of Oriental fruit fly. Detailed description regarding the design and implementation of components utilized in the monitoring system is given in the following subsections. 2.2. Design and implementation The remote agroecological monitoring system built upon WSN and GSM technologies is presented. The proposed monitoring system includes three major parts, a set of wireless multi-sensor nodes (WMNs), a remote sensing information gateway (RSIG), and a host control platform (HCP). The WMN is deployed in a fruit orchard to measure the meteorological parameters (e.g. temperature, humidity, and illumination) and the number of Oriental fruit flies captured by an automatic counting trap attached to the WMN. All WMNs report the sensing information to the RSIG via ZigBee, which is one of the low-power and short-range communication technologies that have been widely used for WSNs. The sensing information acquired from all WMNs is organized into standard format for Short Message Service (SMS), and then send the messages to the HCP via GSM platform. The HCP stores all sensing information in an open database available for public access. The overall conceptual block diagram of the proposed monitoring system, which is a mechatronics application that integrates technologies in mechanics, electronics, communications, and information
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
3
Fig. 1. The proposed monitoring system. (a) Overall conceptual block diagram of the system, and (b) Integration architecture of the proposed system that integrates technologies in mechanics, electronics, communications, and information engineering.
engineering, is depicted in Fig. 1. Next, we describe the design principles of the components that are used in the proposed remote agroecological monitoring system. 2.2.1. Mechanics: automatic counting trap in WMNs Traditionally, field researchers placed a mixture of chemical attractant (i.e., methyl eugenol) and insecticide in a trap to control and suppress the population size of male Oriental fruit fly. In this study, a double-counting mechanism is integrated to form an automatic counting trap attached to the WMNs (Jiang et al., 2008), and the design of the trap pathway is as depicted in Fig. 2a. The number of trapped flies is counted by an 8051 microcontroller as the flies sequentially trigger the infrared photo-interrupters, as shown in Fig. 2b. If the fly crawls back and forth in the trap pathway, as the scenarios depicted in Fig. 2c, the fly is then excluded by the double-counting mechanism (Jiang et al., 2008). In order to simplify the design of the automatic counting trap, this study ignores the following event since the chance of occurrence is quite rare – the fly captured by the trap survives the toxicities with the poisoned attractant, crawls back to the entrance of the pathway, enters the trap again, and consecutively pass through the infrared photo-interrupters, as depicted in Fig. 2d. In addition, the design of the pathway is to reduce the occurrence of such condition.
2.2.2. Electronics: multi-sensor and control circuits in WMNs The WMN consists of an automatic counting trap and a wireless sensor node (Octopus II (Sheu et al., 2008)). The external configuration and internal architecture of a WMN are shown in Fig. 3. The Octopus II is built upon an MSP430 microcontroller made by Texas Instruments, Inc. The wireless sensor node is equipped with a temperature sensor, a humidity sensor, and an illumination sensor. These sensors have been calibrated with high precision meters using linear regression calibration. These sensors are integrated on a compact sensor board that its measurement is time-synchronized with the number of trapped Oriental fruit flies determined by the automatic counting trap. All sensing information is organized into standard format for ZigBee protocol, and sends the data via CC2420, a radio transceiver that follows the IEEE 802.15.4 specification with 250 kbps peak transmission rate. Given that this agroecological monitoring system is applied to the field, wireless communication is easily influenced by terrain and barriers that may cause communication interruption. In order to increase the stability of communication signals, an additional antenna is attached to the system. The WMN has a hybrid power supply that integrates a battery (battery voltage is 12 V with energy storage equals to 36 amp h) and a solar power panel (power rating is 20 W).
4
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
Fig. 2. Illustrations of the trap implementation and double-counting mechanism. (a) Design of the trap pathway integrated with the automatic pest counting trap, and the close-up view of the pathway. (b) Scenario that the flies sequentially trigger the infrared photo-interrupters, and the 8051 microcontroller adds one to the final count, (c) and (d) show the scenarios that the fly crawls back and forth in the trap pathway, and such events are excluded by the double-counting mechanism, and (e) depicts the scenario that the fly captured by the trap crawls back to the entrance of the pathway, and then enters the trap again. The event in (e) is ignored since the chance of occurrence is rare.
Fig. 3. Internal configuration and external architecture of a WMN.
Fig. 4. External architecture and internal configuration of a RSIG.
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
2.2.3. Communications: interface between ZigBee and GSM The RSIG is constructed using an industrial personal computer, a ZigBee transmission module, a GSM module, and a weather station that can be used to obtain a variety of environmental readings. All components are placed in a Stevenson house, and the external configuration and internal architecture of an RSIG are shown in Fig. 4. The software of the RSIG is designed to control the ZigBee transmission module and the GSM module. Therefore, the RSIG is the communication interface between ZigBee and GSM platforms, such that the WMNs would be able to send sensing information to the HCP, and the HCP is then able to issue commands to control the operation of the remote agroecological monitoring system. The high-precision weather station, WS-2310 produced by LaCrosse Technology, Inc.), installed on the RSIG includes sensors to measure local temperature, humidity, wind speed, wind direction, atmospheric pressure, wind chill, dew point temperature, and rainfall volume. The meteorological information acquired from the weather station is sent to the HCP as reference. First, the RSIG checks all input/output interfaces and starts a topology initialization process to establish initial linkage tables for all WMNs. During the topology initialization process, the initial link tables are transmitted to the HCP via the GSM module. The HCP calculates the optimal routing paths for all WMNs, and send them back to the RSIG for setting up optimal routing tables for all WMNs. The RSIG collects the sensing data from the WMNs and itself every 30 min. Finally, the HCP generates a number of SMS text messages, and sends them to HCP over a GSM network. 2.2.4. Information engineering: database and inquiry interface The HCP is built using a computer and a GSM module. After the RSIG send the sensing information to the HCP via GSM platform, the HCP stores the data in a MySQL database for further analysis. A graphical user interface (GUI) is developed in LabVIEW to provide functions of – user authentication, GSM module control, optimizing routing paths for WMNs (Chen et al., 2009; Jiang et al., 2009; Jiang et al., 2011), and MySQL inquiry. Furthermore, the HCP is able to monitor the population dynamics of the Oriental
5
fruit fly. If the number of the Oriental fruit flies significantly increases, a short message (SM) alarm will be sent to the administrators or government officials over a GSM network. Such an SM alert is an essential notification for them to initiate pest control strategies before pest outbreaks. A web service interface, as shown in Fig. 5, is established to provide public network access for external users to retrieve real-time and historical records, i.e., fly counts and environmental factors, collected by the proposed WSN–GSM based remote agroecological monitoring system via the Internet. The website is fully integrated with Google Maps and MySQL database to provide efficient data retrieval service on all geographic information collected by the WMNs. 3. Autonomous event detection algorithm 3.1. Events, input vector, and its pre-processing In order to allow the proposed remote agroecological monitoring system to detect transient events, an autonomous event detection algorithm is developed. Except normal status, there are two abnormal transient events to be addressed in this paper, which are (1) sensor fault and (2) pest outbreaks, which are defined as follows. 3.2. Definition of sensor fault (SF) event There is one or more parameters in the sensory data (i.e., temperature, humidity, illumination, and pest count) acquired from a WMN becomes abnormally large or small. 3.3. Definition of pest outbreaks (PO) event The number of pests captured by the automatic pest counting trap in a WMN is significantly higher than that observed in normal days.
Fig. 5. Graphical user interface of the web service interface which is available for external users to retrieve real-time and historical records via the Internet.
6
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
3.4. Definition of normal status (NS) The remaining conditions are defined as normal status (NS). As mentioned above, the proposed autonomous event detection algorithm takes the parameters in the sensing data as its input vector, which is defined as
xk;t ¼ ½T k;t ; Hk;t ; Ik;t ; P k;t
ð1Þ
where xk,t is the input vector acquired from the kth WMN at time t, and xk,t contains four elements, Tk,t, Hk,t, Ik,t, and Pk,t, which are temperature, relative humidity, illumination, and pest counts measured by the kth WMN at time t, respectively. In the proposed system, the dynamic ranges of temperature Tk,t and relative humidity Hk,t are ranged from 5 to 50 (°C) and 35 to 95 (%RH), respectively. The value of illumination Ik,t, can vary from 0 to 80,000 (Lux) or more, we apply logarithm to the illumination Ik,t, so that we can transform Ik,t, to the brightness Bk,t (cd/m2), of the light by
Bk;t ¼ logðIk;t Þ
ð2Þ
According to the regulation of 10-days agroecological bulletin in Taiwan (Agricultural Research Institute, 2011), if the number of Oriental fruit flies captured by a bait trap exceeds 64, 256, and 1024 per 10 days, the government will issue yellow, blue, and red flag alert, respectively. Among these alerts, the red flag alert is a high vigilance alert for Oriental fruit fly threat that require an immediate eradication treatment (e.g. pesticide or protein bait) to suppress the population of Oriental fruit fly. Since the WMNs reports all sensing information every 30 min, and Oriental fruit flies are most active during the day, 4.27 (1024/(10 days12 h2 sampling per hour)) flies captured by a bait trap per 30 min would be the reasonable threshold to issue a red flag alert. Since the value of the threshold is relatively small due to high temporal resolution sampling, it is tough for any computational algorithms to correctly classify the PO and NS events. By transforming Pk,t to P2k;t , it would be easier to distinguish the differences between the PO and NS events since the square function will cause the value of P 2k;t to become more extreme when the value of Pk,t increases. Furthermore, if we transform the value of P 2k;t (in our historical database) into a probability distribution, the mean of P2k;t is 4.785, and the standard deviation is 4.9656. In such a probability distribution, the value of P2k;t must be greater than 16.3368 to be a significantly large value at the 0.01 level of significance. It implies that the value of Pk,t will then become an significant one if it is greater than (16.3368)0.5 = 4.0419, which is close to the threshold of the red flag alert. Therefore, the analytical results show that the transforming Pk,t to P 2k;t is able to assist the proposed system to distinguish PO events from NS events. According to above discussion, the squared ~ k;t can be repest count is used, and the transformed input vector x expressed as below
~ k;t ¼ ½T k;t ; Hk;t ; Bk;t ; P2k;t x
sensing data to a low dimensional feature space. In addition, the neurons around the winner neuron can be updated; therefore, the neurons in the output layer can be arranged in one- or twodimensional space based on the feature contained in the input layer. After the training process is completed, the neurons in the output layer can be visualized via a topology graph. All of the samples are clustered into several groups (three groups in this study) whereas the neurons in each group are close (highly associated) to each other. Furthermore, SOMs is an unsupervised learning method, this characteristic would help the proposed system to distinguish PO events from NS events without knowing the threshold value to issue a red flag alert, such that the results of machine learning would not be affected by human labeling errors. Secondly, SVM is integrated into the proposed method because it is able to determine the nonlinear boundaries between grouped neurons in the map produced by SOMs. After that, the all of the samples are then classified into their corresponding classes, PO, SF, and NS events. In the following subsections, the theoretical foundations of SOMs and SVM used in this study are briefly introduced.
ð3Þ
Because different orchards located in different areas commonly have dissimilar ecological models, detecting ecological events by setting an empirical threshold to the entire system can cause a large amount of false-positives (or say, false-alarms). Therefore, in ecology, the proposed system would become an impractical one if we set an empirical threshold to detect ecological events in wild orchards. To overcome this problem, it is necessary to use computational methods to effectively detect the ecological events that occurred in the fruit orchards. In this study, the autonomous event detection algorithm is constructed by combining two algorithms, self-organizing maps (SOMs) and support vector machine (SVM). SOMs is used because it is a method which is capable of mapping high dimensional
3.5. Self-Organizing Maps (SOMs) Self-Organizing Maps (SOMs) (Kohonen, 1989) are unsupervised networks widely used in both data analysis and vector quantization because they transform the information from original input space (n dimensional) into a reduced output space (two dimensional) while preserving the most important topological and distribution relation of the original input vector. Such characteristics of SOMs are used in this study for analyzing the sensing information received from the WMNs via WSN and GSM platforms. The main purpose of SOMs is to define a mapping from the input ~k;t 2 R4 onto a two dimensional array of neurons. vector x The number of neurons can vary from a few dozen up to several thousands. In this study, we determine a suitable number of neurons according the size of training dataset. Each neuron is represented by a four dimensional weight vector, and then the weight vector of the nth neuron can be expressed as
mn ¼ ½mT;n ; mH;n ; mB;n ; mP;n
ð4Þ
where mn is the weight vector of the nth neuron, and mT,n, mH,n, mB,n, and mP,n represent the weighting elements for temperature, relative humidity, brightness, and pest counts, respectively. Each neuron is directly connected to neighboring neurons, and the links between neurons indicate the topology or structure of the map. In this study, the sequential training algorithm is chosen to train SOMs, iteratively. In each training epoch, we randomly ~ k;t ) from choose a training sample (the transformed input vector x ~ k;t the training dataset, and we measure the distance between x and all the weight vectors using Euclidean matric
~k;t mn k dx;n ¼ kx
ð5Þ
where dx,n is the Euclidean distance between the transformed input vector and the weight vector of the nth neuron, and |||| is the Euclidean distance measure. The neuron whose weight vector is ~k;t is the best-matching closest to the transformed input vector x unit, which is denoted by c
c ¼ arg mindx;n
ð6Þ
n
When a best-matching neuron c is determined, the weight vectors of the neuron and its topological neighbors on the map are simultaneously updated to move closer to the transformed input ~k;t . The SOM update rule for the weight vector of the nth vector x neuron is
~k;t mn ðjÞ mn ðj þ 1Þ ¼ mn ðjÞ þ aðjÞhc;n ðjÞ½x
ð7Þ
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
7
where j denotes the training step, a(j) is the learning rate at the jth training step, hc,n(j) represents the neighborhood kernel for the neuron n around the best-matching unit (BMU) c at the jth training step. The neighborhood kernel hc,n(j) represents the region and strength of influence that the transformed input vector has on the SOMs, which can be formulated using a Gaussian function as
krn rc k hc;n ðjÞ ¼ exp 2r2 ðjÞ
ð8Þ
where rn and rc are the coordinate of neuron n and the best-matching unit c in R2, respectively. ||rn rc|| is the distance between the best-matching unit c and the neuron n within the output space (map), and r(j) defines the width of the neighborhood kernel at the jth training step. By using the aforementioned training procedures, the neurons in the region around the best-matching neurons are stretched toward the inputted training samples. Finally, after the training epoch reaches j, the training process is terminated. In the trained SOMs, the neurons on the map become ordered, and the neighboring neurons would have similar weight vectors. Assessing the quality of training through the graphical representation of all neurons on the map is possible by checking whether training dataset acquired from the WMNs can be well grouped according to its corresponding events, i.e., NS, SF, and PO. 3.6. Boundary determination via support vector machine (SVM) There are two possible approaches to correctly identify the groups (or say, event groups) on the map, one is visual inspection, and the other is to utilize an automatic boundary determination method. In order to automize the discrimination process on the event groups presented on the map, the support vector machine (SVM) is used to automatically determine the nonlinear boundaries between event groups. It is given that a set of prior knowledge vectors [rn, yn] that contains a transformed input vector and an event label associated with the input vector (coded with ‘‘+1’’ and ‘‘1’’), respectively. The SVM regression function can be formulated by
fj ðrn Þ ¼ wTj rn þ bj
ð9Þ
where j = 1, 2, . . ., N, N denotes the number of hyperplanes that divide the maps into various portions, and wj and bj are the weight vector and the bias term of the jth hyperplane, respectively. The final decision function to determine the event type of rn can be ex^n = sgn[fj(rn)]. bj/||wj|| is the distance between the jth pressed as y hyperplane and the origin where ||wj|| is the Euclidean norm of wj. The distance between the closest positive and negative samples is equal to 2/||wj||. Consequently, the parameters of the jth hyperplane are optimized by maximizing this margin subject to the constraints defined by
hwTj
rn i þ bj P 1 for yn ¼ 1 8n;
ð10Þ
During the optimization process, the transformed input vectors may be non-separable. Therefore, a positive slack variable nn is introduced to measure the misclassification error for the coordinate of neuron n, as well as to deal with noisy and nonlinearly separable data. The optimization problem can be expressed as
minimize J ¼ w;n
X 1 kwj k2 þ C nn 2 n
ð11Þ
under the modified constrain 8n; yn fj ðrn Þ P 1 nn , and the positive parameter C is the regularization parameter that represents the relative importance of the misclassification error nn. In this study, we need to divide the maps into three event types (i.e., NS, SF, and PO), that requires two hyperplanes (N = 2) to set the boundaries between {NS + PO, SF} and {NS, PO + SF}, as illustrated in Fig. 6. We
Fig. 6. Results of SVM boundary determination for dividing the map into three event groups.
can see that sensor fault event can be distinguished because the sensor readings are significantly different from normal readings, as shown in Fig. 7. It is tougher to determine the difference between NS and PO events. The interest of this study is to determine the nonlinear boundaries between the event groups, it is necessary to map the coordinate of neuron rn from the input space R2 onto a feature space with higher dimension (i.e., a Hilbert space H) using a nonlinear transfer function u. With such a transformation, SVM can find a set of linear hyperplanes to divide the maps into three event types. By using a commonly-used polynomial kernel K(rn,1, rn,2) = u(rn), the discrimination function of the jth hyperplane can be re-expressed by
fj ðrn Þ ¼ wTj uðrÞ þ bj
ð12Þ
where
pffiffiffi
uðrÞ ¼ ðr2n;1 ; 2rn;1 rn;2 ; r2n;2 ÞT
ð13Þ
The SVM optimization problem can be solved according to the Lagrangian formulation L by obtaining a dual problem in which the following convex objective function should be maximized as
maximizeL ¼ a
subject to
X 1X ai þ y y ai aj huðri Þ; uðrj Þi 2 ij i j i
8X < yj a i ¼ 0 :
i
ð14Þ
ð15Þ
C P ai P 0 and C > 0:
where ai are positive Lagrange multipliers (which are associated with the coordinate of neuron rn in the training dataset). This problem can be solved by Quadratic Programming method (Bottou et al., 2007) or a more efficient solver, LIBSVM (Joachims, 2006). Using a standard computer with the Intel Core-2 Duo-class processor and MATLAB programming, the training process of an SOM with a size of 8 8 neurons only takes 7.48 s to complete the training, and the boundary determination process of an SVM using the LIBSVM solver can be done in 1.26 s. After the SOM and SVM are trained, the proposed system can provide the result of the event detection process in real-time. The CPU times required to complete the training processes of FCM, KM, NF, and MLP are 5.98, 4.47, 13.35, and 11.27 s, respectively. Similar to the SOM and SVM, they
8
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
Fig. 7. Faulty readings obtained from (a) temperature sensor, (b) humidity sensor, (c) illumination sensor, and (d) pest counting trap. Please note that the data used in (a), (b) and (c) is obtained on August 26, 2011, and the data used in (d) is obtained on October 31, 2011.
can provide the result of the event detection process in real-time, as well. 4. Experimental results 4.1. Deployment of the proposed agroecological monitoring system We deploy the developed remote agroecological monitoring system in an orange orchard located at Yuanshan Township, ILan County, Taiwan (N24°440 03.0600 , E121°400 43.0300 ), to evaluate the event detection results of the proposed method. The
deployment of the monitoring system at the selected orchard is shown in Fig. 8. The devices deployed in the orchard comprise 10 WMNs and an RSIG. The sensing data acquired from WMNs and RSIG is transmitted to an HCP using SMS via GSM platform every 30 min, and the WMNs send the sensing data to the remote sensing information gateway (RSIG) using maximum radio power (1 mW). The minimum communication range of the WMNs in the field is also evaluated, and we can achieve stable wireless transmission between paired WMNs distanced within 30 m. Because the area of the deployment site is around 40 m 60 m, and any two neighboring WMNs are distanced by 3–6 m, we can
Fig. 8. Deployment of the proposed agroecological monitoring system at the orange orchard located at Yuanshan Township, I-Lan County, Taiwan. The WMNs are distributed in the experimental field, and the RSIG also deployed in the orchard to collect sensing data measured by the WMNs.
9
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12 Table 1 Daily operation of the infrared photo-interrupters in each season. Season
Day (activate) Night (deactivate)
Table 2 Monthly availabilities
a
Summer (June–August)
Autumn (September–November)
Winter (December–February)
5 am–6 pm 6 pm–5 am
4 am–8 pm 8 pm–4 am
6 am–6 pm 6 pm–6 am
6:30 am–5:30 pm 5:30 pm–6:30 am
of all WMNs and RSIG.
Period
a
Spring (March–May)
WMN #1 (%)
#2 (%)
#3 (%)
#4 (%)
#5 (%)
#6 (%)
#7 (%)
#8 (%)
#9 (%)
#10 (%)
RSIG (%)
July August September October November December
2010
47 53 37 12 51 84
68 86 36 19 3 79
47 35 59 33 87 70
36 27 30 41 34 62
61 60 7 32 44 51
57 39 35 38 93 86
68 87 57 43 89 86
62 83 55 42 3 67
68 86 57 28 15 72
58 49 32 38 52 22
66 60 54 50 47 30
January February March April May Average
2011
70 40 96 99 95 62
32 26 88 1 49 44
50 30 64 95 79 59
2 8 34 68 79 38
1 1 1 1 49 28
71 40 96 99 95 68
71 34 58 82 82 69
56 26 84 99 95 61
55 44 61 80 81 59
48 40 29 1 55 39
66 42 91 97 92 63
Availability = # received samples/# expected samples.
Table 3 Summary of the samples in the training and test data sets. Training data set
Test data set
Spring
Summer
Autumn
Winter
Spring
Summer
Autumn
Winter
135 130
160 150
409
490
530
471
NS
Day Night
125 125
125 125
PO
Day
100
100
50
10
7
67
14
6
SF
Day Night
25 25
25 25
45 40
40 40
184
43
56
123
then ensure high-quality wireless transmission between WMNs and the RSIG. Furthermore, if a WMN is desynchronized from the network due to some reasons, the WMN will not be able to send sensing data to the RSIG. Under such circumstances, the RSIG would broadcast a re-initialization command to reset the entire network, and the desynchronized WMN is then synchronized and allowed to rejoins the network. Therefore, with this mechanism, the reliability of the network can be guaranteed. Because Oriental fruit flies are inactive during the night, the infrared photo-interrupters in WMNs are switched off (on) at night (day) in order to reduce power consumption. Since the length of day and night varies in each season, the daily operation of the infrared photo-interrupters in each season is arranged as summarized in Table 1. The proposed remote agroecological monitoring system becomes operational on July 1, 2010 till the present time. During the experiment period, we let the system run its course without being influenced by manual assistance, so we can get more diversity in the sensing data sets. Thus, only two maintenance services were arranged. The first service was arranged in September 2010 to repair the damage caused by typhoon Fanapi which devastated eastern coast of Taiwan on September 19, 2010. The second one is a scheduled inspection and repair service that replaces naturally damaged components using spare parts. The monthly availabilities of all WMNs and RSIG are summarized in Table 2. Since the performance of the proposed study depends upon the accuracy of the sensing data, we have conducted a series of analysis while
calibrating the sensing data measured by our system and a weather station. By utilizing linear regression to the data collected by our system and weather station, the average R2 values for temperature, humidity, and illumination sensors are 0.8896, 0.8972, and 0.9069, respectively. The results show strong associations between data obtained from the sensors and weather station. Furthermore, the automatic pest counting trap is also evaluated, and the average pest counting accuracy is 96.3%; thus, the sensor precisions are ensured. 4.2. Performance evaluation of the proposed autonomous event detection algorithm As previously mentioned in Section 3, there are three types of events addressed in this work, NS, SF, and PO. Among all sensing data acquired from the system, we selected 400 (600) samples in each of the season to serve as training (test) data set for the proposed autonomous event detection algorithm. Table 3 summarizes the number of training and the test samples acquired by the proposed monitoring system in each season. Four SOMs of size 8 8 neurons are created to identify the event types of samples in training and test data sets collected in every season. Both SOMs and SVM utilized in the autonomous event detection algorithm are implemented using the neural network toolbox and bioinformatics toolbox in MATLAB. The number of neurons is determined via the ratio of knowledge by unknown
10
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
autumn (CMA), and winter (CMW). The patterns in these models show that the proposed autonomous event detection algorithm successfully discriminates different event groups into separated regions. The performances of the autonomous event detection algorithm in each season were studied using the training data sets collected in every season. We also compared the proposed method to fuzzy c-means (FCM) clustering (Bezdek, 1981), k-means clustering (KM) (MacQueen, 1967), neuro-fuzzy (NF) classifier (Abraham, 2005), and multi-layer perceptron (MLP) classifier (Haykin, 1998). The purpose of the comparison is to show that the detection of agroecological events is not a trivial problem that can be solved simply using a predetermined threshold. The FCM and KM are applied to the training data set, and separate the samples into three event groups. The supervised learning methods, NF and MLP, are implemented with the same ratio of knowledge by unknown variables (6.25). Both methods are applied to classify the samples into three event groups. The sensitivity, specificity, accuracy, and precision yielded by all methods are reported. The definitions of these performance indices are given at below.
Fig. 9. Clustering and discrimination models yielded by the proposed autonomous event detection algorithm for (a) spring, (b) summer, (c) autumn, and (d) winter.
variables (400 samples 4 parameters per sample/(64 neurons 4 parameters per weight vector) = 6.25, which is a reasonable ratio). After the training process of SOMs is completed, SVM is applied to the map in follow to discriminate three event groups using two hyperplanes. The clustering and discrimination results of different seasons over a year are demonstrated in Fig. 9 where each result represents the trained model for its corresponding season, namely clustering model for spring (CMSP), summer (CMS),
Sensitivity ¼ TP=ðTP þ FNÞ;
ð16Þ
Specificity ¼ TN=ðTN þ FPÞ;
ð17Þ
accuracy ¼ ðTP þ TNÞ=#samples and;
ð18Þ
precision ¼ TP=ðTP þ FPÞ:
ð19Þ
where TP, TN, TP, and FP are numbers of true-positives, true-negatives, false-positives, and false-negatives, respectively. The experimental results show that the proposed autonomous method yielded the averaged sensitivity, specificity, accuracy, and precision of 98%, 100%, 100%, and 100%, respectively, whereas those of FCM are 82%, 82%, 81%, and 54%, respectively. KM performs poorly as expected. The performances yielded by NF and MLP are the same (quite close) after training, but their
Table 4 Summary of the samples in the training and test data set. Season
Method
NS v.s. PO Performance Indices TP
Spring
Summer
Autumn
Winter
TN
FP
FN
NS v.s. SF Performance Indices Sensitivity (%)
SOM/ SVM FCM KM NF MLP
6
409
0
1
85.71
5 5 5 5
75 73 284 269
334 336 125 140
2 2 2 2
71.43 71.43 71.43 71.43
SOM/ SVM FCM KM NF MLP
67
490
0
0
30 21 53 50
365 366 464 439
125 124 26 51
37 46 14 17
14
530
0
0
6 4 11 10
420 417 284 269
110 113 246 261
8 10 3 4
6
471
0
0
100
6 6 6 6
391 51 416 403
80 420 55 68
0 0 0 0
100 100 100 100
SOM/ SVM FCM KM NF MLP SOM/ SVM FCM KM NF MLP
100 44.78 31.34 79.10 74.63 100 42.86 28.57 78.57 71.43
Specificity (%) 100 18.29 17.80 69.51 65.85 100 74.49 74.69 94.69 89.59 100 79.25 78.68 53.58 50.75 100 83.01 10.83 88.32 85.56
Accuracy (%) 99.76 19.18 18.71 69.54 65.95 100 70.92 69.48 92.82 87.79 100 78.31 77.39 54.23 51.29 100 83.23 11.95 88.47 85.74
Precision (%)
TP
TN
FP
FN
Sensitivity (%)
Specificity (%)
Accuracy (%)
Precision (%)
100
184
409
0
0
100
100
100
100
184 184 184 184
409 409 409 409
0 0 0 0
0 0 0 0
100 100 100 100
100 100 100 100
100 100 100 100
100 100 100 100
43
490
0
0
100
100
100
100
43 43 43 43
490 490 490 490
0 0 0 0
0 0 0 0
100 100 100 100
100 100 100 100
100 100 100 100
100 100 100 100
56
530
0
0
100
100
100
100
56 56 56 56
530 530 530 530
0 0 0 0
0 0 0 0
100 100 100 100
100 100 100 100
100 100 100 100
100 100 100 100
123
471
0
0
100
100
100
100
123 123 123 123
471 471 471 471
0 0 0 0
0 0 0 0
100 100 100 100
100 100 100 100
100 100 100 100
100 100 100 100
1.47 1.46 3.85 3.45 100 19.35 14.48 67.09 49.50 100 5.17 3.42 4.28 3.69 100 6.98 1.41 9.84 8.11
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
11
Fig. 10. Demonstration of the alert message function provided by the proposed system when (a) an PO or (b) SF event has occurred.
performances still fall behind the proposed method, whereas the averaged sensitivity, specificity, accuracy, and precision yielded by NF (MLP) are 91% (90%), 88% (86%), 88% (86%), and 61% (58%), respectively. Comparison of the proposed autonomous event detection algorithm to the other methods is summarized in Table 4, the experimental results show that: (1) all of the SF events are correctly identified by the proposed method and other alternative methods (i.e., FCM, KM, NF, and MLP); (2) the majority of the PO events in all seasons (99%) are classified by the proposed method with promising performance, and show that the proposed method significantly outperforms the rest of the alternative methods; (3) by combining SOMs and SVM, the proposed monitoring system provides a promising way to detect agroecological events in open-air farms. Furthermore, other alternative methods are overcome by the proposed one because they are linear clustering models (e.g. FCM or KM) that are unable to separate nonlinear agroecological data, or they are supervised learning methods (e.g. NF or MLP classifiers) that the training results are often difficult to be generalized. As mentioned in previous section, when the proposed autonomous event detection algorithm identifies an SF or PO event, the monitoring system will send an alert message to the system administrators’ or government officials’ cell phone via GSM network. 4.3. Alert message for PO event ‘‘PO event has occurred at WMN #5 in I-Lan (5 OFFs captured half an hour ago).’’ 4.4. Alert message for SF event ‘‘SF event has occurred at WMN #3 in I-Lan (half an hour ago).’’ Please note that ‘‘#5’’ and ‘‘#3’’ in the messages in above are for demonstration only, and the numbers are changeable depends on the property of the event. The function of sending alert messages is also examined using sensing data acquired in real-world scenario. The experimental results are demonstrated in Fig. 10. It shows that, by combining the remote agroecological monitoring system and the autonomous event detection algorithm, the government officials in agricultural authorities would be able to react to the agroecological events via the alert messages sent by the monitoring system presented in this study.
5. Conclusion In this paper, we have presented a remote agroecological monitoring system and an autonomous event detection algorithm for Oriental fruit fly, Bactrocera dorsalis (Hendel). The remote agroecological monitoring system is built upon two different wireless communication protocols, ZigBee and GSM, and three major components, wireless monitoring nodes (WMNs), a remote sensing information gateway (RSIG), and a host control platform (HCP). These components are deeply integrated in mechatronics techniques, including mechanics, electronics, communication, and information engineering. WMNs transmit the sensing data (temperature, relative humidity, illumination, and the number of Oriental fruit flies captured by the automatic pest counting trap) to the RSIG, and the RSIG relay the data to the database server in the HCP for storage and analysis (real-time sensing data is available at URL: http://140.112.94.60/wsn3/en/WebForm1.aspx). The autonomous event detection algorithm is a self-organizing maps (SOMs)-based approach that is able to group the sensing data into three event types, normal status (NS) event, pest outbreaks (PO) event, and sensor fault (SF) event. By utilizing support vector machine (SVM), the discrimination process on the event groups presented on the map can be further automized. Finally, the proposed system is deployed in an orange orchard in I-Lan, Taiwan for performance evaluation in real-world scenarios. During the experiment period, the system runs without much manual assistance in order to get more diversity in the sensing data sets. Comparing with other existing methods used in other applications, i.e., fuzzy c-means, k-means, neuro-fuzzy and multilayer perceptron, the experimental result of the proposed autonomous event detection algorithm outperforms those existing methods, and is relatively efficient and effective for the sensor fault and pest outbreaks detection system. The proposed study also offers a real-time warning system to inform system administrators and government officials the occurrence of crucial events (e.g. PO and SF events) via GSM platform, so that the farms and future food security can be protected. Based on the sensing data collected by the proposed system, the population dynamics and population forecast models of the Oriental fruit fly can be built. In addition, after the ecologists in the government of Taiwan finish the development of a generalized delta-based ecological model, the proposed system can also be used to issue a warning of PO event when the pest population rapidly changes. We leave these research tasks as our future works.
12
M.-S. Liao et al. / Computers and Electronics in Agriculture 88 (2012) 1–12
Acknowledgements This work was supported in part by the National Science Council of the Executive Yuan and the Council of Agriculture of the Executive Yuan of Republic of China under contracts NSC 982218-E-002-039, NSC 99-2218-E-002-015, 98AS-6.1.4-FD-Z1, 99AS-6.1.5-FD-Z1, and 100AS-6.1.2-BQ-B2. This work was also supported by National Science Council, National Taiwan University and Intel Corporation under Grants NSC99-2911-I-002-201, NSC 100-2911-I-002-001, and 10R70501. References Abraham, A., 2005. Adaptation of fuzzy inference system using neural learning, fuzzy system engineering: theory and practice. In: Nedjah, N., et al. (Eds.), Studies in Fuzziness and Soft Computing. Springer Verlag, Germany, pp. 53–83 (Chapter 3). Agricultural Research Institute, 2010a. Ten-Day Bulletin of Essential Insect Pests of Vegetables and Fruits. Council of Agriculture, Executive Yuan, Taiwan (in Chinese).
(accessed 07.10.10). Agricultural Research Institute, 2010b. Ten-Day Bulletin of Essential Insect Pests of Vegetables and Fruits. Council of Agriculture, Executive Yuan, Taiwan (in Chinese). (accessed 15.06.11). Armstrong, K.F., Carmichael, A.E., Milne, J.R., Raghu, S.R., Roderick, G.K., Yeates, D.K., 2004. Invasive phytophagous pests arising through a recent tropical evolutionary radiation: the Bactrocera dorsalis complex of tropical fruit flies. Annu. Rev. Entomol. 50, 293–319. Bezdek, J.C., Rajasegarar, S., Moshtaghi, M., Leckie, C., Palaniswami, M., 2011. Anomaly detection in environmental monitoring networks. IEEE Comput. Intell. Mag. 6 (2), 52–58. Bezdek, J.C., 1981. Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press, New York, USA. Bottou, L., Chapelle, O., DeCoste, D., Weston, J., 2007. Large Scale Kernel Machines. MIT Press, Massachusetts, USA. Chen, C.C., McCarl, B.A., 2001. An investigation of the relationship between pesticide usage and climate change. Climatic Change 50 (4), 475–487. Chen, P., Ye, H., Liu, J., 2006. Population dynamics of Bactrocera dorsalis (Diptera: Tephritidae) and analysis of the factors influencing the population in Ruili, Yunnan Province, China. Acta Ecol. Sin. 26 (9), 2801–2809. Chen, C.P., Chuang, C.L., Tseng, C.L., Yang, E.C., Jiang, J.A., 2009. A novel energyefficient adaptive routing protocol for wireless sensor networks. J. Chin. Soc. Mech. Eng. 30 (1), 59–65. Cheng, E.T., Hwang, Y.B., Kao, C.H., Chaing, M.Y., 2002. An area-wide control program for the Oriental fruit fly in Taiwan. In: Proc. Symp. Insect Ecology and Fruit Fly Management, pp. 57–71. Dissanayake, M., Newman, P., Clark, S., Durrant-Whyte, H., Csorba, M., 2001. A solution to the simultaneous localization and map building (SLAM) problem. IEEE Trans. Robot. Autom. 17 (3), 229–241. Drew, R.A.I., Raghu, S., 2002. The fruit fly fauna (Diptera: Tephritidae: Dacinae) of The rainforest habitat of the western ghats, India. Raffles Bull. Zool. 50 (2), 327– 352. Fischer, G., Shah, M., van Velthuizen, G., 2002. Climate Change and Agriculture Vulnerability. International Institute for Applied Systems Analysis, Laxenburg, Astria. Haykin, S., 1998. Neural Networks: A Comprehensive Foundation, second ed. Prentice Hall, London, UK.
Huang, J., Farritor, S.M., Qadi, A., Goddard, S., 2006. Localization and follow-theleader control of a heterogeneous group of mobile robots. IEEE/ASME Trans. Mechatron. 11 (2), 205–215. Hung, Y.T., Tsai, W.H., Kuo, K.C., 2008. Oriental fruit fly management in Taiwan: current and future. In: Proc. the International Symposium on the Recent Progress of Tephritid Fruit Flies Management, pp. 5–9. Jiang, J.A., Tseng, C.L., Lu, F.M., Yang, E.C., Wu, Z.S., Chen, C.P., Lin, S.H., Lin, K.C., Liao, C.S., 2008. A GSM-based remote wireless automatic monitoring system for field information: a case study for ecological monitoring of the Oriental fruit fly, Bactrocera dorsalis (Hendel). Comput. Electron. Agric. 62 (2), 243–259. Jiang, J.A., Chen, C.P., Chuang, C.L., Lin, T.S., Tseng, C.L., Yang, E.C., Wang, Y.C., 2009. CoCMA: energy-efficient coverage control in cluster-based wireless sensor networks using a memetic algorithm. Sensors 9 (6), 4918–4940. Jiang, J.A., Lin, T.S., Chuang, C.L., Chen, C.P., Sun, C.H., Juang, Y.J., Lin, J.C., Liang, W.W., 2011. A QoS-guaranteed coverage precedence routing algorithm for wireless sensor networks. Sensors 11 (4), 3418–3438. Joachims, T., 2006. Training linear SVMs in linear time. In: Proc. ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 217– 226. Kohonen, T., 1989. Self-Organisation and Associative Memory, third ed. SpringerVerlag, Berlin, Germany. Lin, M.Y., Chen, S.K., Liu, Y.C., Yang, J.T., 2005. Pictorial key to 6 common species of the genus Bactrocera from Taiwan. Plant Prot. Bull. 47, 39–46. MacQueen, J.B., 1967. Some methods for classification and analysis of multivariate observations. In: Proc. 5th Berkeley Symposium on Mathematical Statistics and Probability, pp. 281–297. Malacrida, A., Gomulski, L., Bonizzoni, M., Bertin, S., Gasperi, G., Guglielmino, C.R., 2007. Globalization and fruitfly invasion and expansion: the medfly paradigm. Genetica 131, 1–9. Metcalf, R.L., Metcalf, E.R., 1992. Fruit flies of the family Tephritidae. In: Metcalf, R.L., Metcalf, E.R. (Eds.), Plant Kairomones in Insect Ecology and Control. Chapman & Hall, New York, USA, pp. 109–152. Rosenzweig, C., Iglesias, A., Yang, X.B., Epstein, P.R., Chivian, E., 2001. Climate change and extreme weather events; implications for food production, plant diseases, and pests. Global Changes Hum. Health 2 (2), 90–104. Sheu, J.P., Chang, C.J., Sun, C.Y., Hu, W.K., 2008. WSNTB: A testbed for heterogeneous wireless sensor networks. In: Proc. 1st IEEE International Conference on UbiMedia Computing, Lanzhou, Gansu, China, pp. 338–343. Smith, P. H., 1989. Behavioral partitioning of the day and circadian rhythmicity. In: Robinson, A.S., Hooper, G. (Eds.), Fruit Flies: Their Biology, Natural Enemies, and Control (World Crop Pests Series), vol. 3B. Elsevier, Amsterdam, NL, pp. 325– 341. Subramaniam, S., Palpanas, T., Papadopoulos, D., Kalogerakiand, V., Gunopulos, D., 2006. Online Outlier Detection in Sensor Data Using Nonparametric Models, J. Very Large Data Bases, VLDB 2006. Sutherst, R.W., Constable, F., Finlay, K.J., Harrington, R., Luck, J., Zalucki, M.P., 2011. Adapting to crop pest and pathogen risks under a changing climate. Wiley Interdiscipl. Rev.: Clim. Change 2 (2), 220–237. Valadão, H., Hay, J., Tidon, R., 2010. Temporal dynamics and resource availability for drosophilid fruit flies (Insecta, Diptera) in a gallery forest in the Brazilian Savanna. Int. J. Ecol. 2010, 152437. Vargas, R.I., Miyashita, O., Nishida, T., 1984. Life history and demographic parameters of three laboratory-reared tephritids (Diptera: Tephritidae). Ann. Entomol. Soc. Am. 77, 651–656. Weems, H.V., Heppner, J.B., Nation, J.L., Fasulo, T.R., 2008. Oriental Fruit Fly, Bactrocera dorsalis (Hendel) (Insecta: Diptera: Tephritidae). Gainesville: Institute of Food and Agricultural Sciences, University of Florida, FL, Tech. Rep. EENY-083. Zhang, Y.Y., Chao, H.C., Chen, M., Shu, L., Park, C.H., Park, M.S., 2010. Outlier detection and countermeasure for hierarchical wireless sensor networks. IET Inf. Secur. 4 (4), 361–373.