Neural Computing and Applications https://doi.org/10.1007/s00521-018-3386-4
(0123456789().,-volV)(0123456789().,-volV)
ORIGINAL ARTICLE
Smart heterogeneous precision agriculture using wireless sensor network based on extended Kalman filter Yousef E. M. Hamouda1 • Mohammed M. Msallam1 Received: 11 August 2017 / Accepted: 21 February 2018 Ó The Natural Computing Applications Forum 2018
Abstract Smart heterogeneous precision agriculture (SHPA) using a wireless sensor network is introduced for limited energy sensor nodes, deployed in an agricultural farm which is divided into a number of heterogeneous agricultural areas. At each time step, the extended Kalman filter is adopted to measure and to predict the agricultural parameters including the soil moisture and temperature so that the noise associated with noisy measurements is filtered. After that, sensor node selection algorithm proactively selects the sensor nodes for each individual area to sense the agricultural parameters. Thus, the network lifetime and sensing accuracy are improved. The sampling interval for each crop is predefined based on the crop types and agricultural requirements so that the crop yields are improved. Also, the design, framework, algorithms, and architecture of SHPA are considered and proposed. Compared with other schemes, simulation results show that the proposed SHPA scheme eliminates the noise associated with the measurements, improves the network lifetime and sensing accuracy, and enhances the crop yields. Keywords Extended Kalman filter Sensor node selection Precision agriculture Wireless sensor networks
1 Introduction The rapid progress in technology and scientific research across various fields of industry, especially in medicine, has contributed in an exponential population growth. Thus, the demands on water and agricultural production have risen. In fact, 64% of the available land is already employed for agriculture using 85% of the available freshwater in the world [1]. Moreover, agriculture will use more water due to the increasing population and food demands. Unfortunately, 40% of agricultural water in developing countries is wasted due to evaporation, spills, and absorption inside the deep soil layer [2]. Furthermore, the availability of water resources including groundwater is reducing because of environmental and climate changes. This continuous decreasing of groundwater will affect & Yousef E. M. Hamouda
[email protected] Mohammed M. Msallam
[email protected] 1
Department Computer Science, Faculty of Applies Sciences, Al-Aqsa University, P.O. Box: 4051, Gaza, Palestine
seriously the farmers [3]. Hence, research has focused on developing and innovating within agriculture and farming to overcome the limitations of traditional farming and to improve the crop production and water conservation. Wireless sensor networks (WSNs) have become an emerging phenomenon in industry, for civil and military purposes. WSNs can provide virtual snapshots of the physical world by interpreting the physical events. They typically consist of hundreds or thousands of tiny sensor nodes, connected to each other via wireless communication protocols [4, 5]. WSNs have many advantages, such as easy random deployment, low cost, and small size. The sensor nodes cooperate [6] to sense, compute, and transmit information from harsh physical environments to an external base station or sink, which is responsible for forwarding the desired information from the WSN to the headquarters. Sensor nodes are scattered in the sensor field, which is the area in which the sensor nodes are deployed [7]. Deployment of the sensor nodes can be either random, such as in a disaster situation where, for example, sensor nodes are dropped from an airplane [8], or in planned manner, such as the deployment of WSNs in smart homes or for fire alarm systems. Most WSNs have fixed sensor
123
Neural Computing and Applications
nodes. However, the mobility of sensor nodes is desirable in many applications [6]. Nowadays, WSNs are employed for a wide range of applications within commercial and industrial fields, such as health care, target tracking, health monitoring, weather monitoring, pollution monitoring, environmental monitoring, structural monitoring, machine condition monitoring, wildlife habitat monitoring, navigation, forest fire monitoring, manufacturing job flows, control of moving vehicles, smart homes, inventory control, decision support systems, disaster management, surveillance of people or vehicles, biochemical material detection, and agriculture. Also, WSNs can be used in military applications, including target field imaging, intrusion detection, enemy vehicle tracking, detecting illegal crossings, and security/tactical surveillance [9]. One of the most crucial needs of human life is agriculture as it provides food supplier for humans and animals, and it benefits human employment and national economy. WSNs recently have been employed in modern agriculture and farming. Precision agriculture (PA) or precision farming (PF) refers to automation in agriculture using information technology. PA is defined as the techniques of applying PA inputs (i.e., farming parameters and resources) at the right location and time to optimize farming production, improve resource utilization, provide online monitoring for farmers, reduce the agriculture costs, and reduce the human effort, subject to minimizing the environmental impact [10–12]. Farming parameters and resources include site-specific application of water, fertilizer doses, pesticide, herbicides, and monitoring soil moisture and air temperature [13]. Consequently, PA is an agriculture system based on information and technology. Therefore, PA needs decision support systems to achieve its goals. Thus, WSNs are employed in PA to monitor, optimize, and measure different farming and sowing parameters and resources. These parameters are transmitted wirelessly to farmer so that appropriate actions can be taken [14, 15]. Agricultural parameters, such as soil moisture, can be measured by using either direct sensing or remote sensing. Direct sensing employs a large number of sensor nodes, placed directly on the land, while remote sensing is performed using aircraft and satellites, based on electromagnetic wave reflection and absorption by the soil or imagery and maps [16, 17]. However, direct sensing can collect data that give more accurate measurements with better resolution and timing, and lower cost, compared with remote sensing. In addition, direct sensing is not affected by the weather and field conditions [18]. One of the PA applications is irrigation, which is the process of delivering water to soil. Water requirements of plants vary according to crop species, crop growth stage, soil moisture, soil temperature, and soil type. Irrigation management aims to optimally irrigate the crops as needed.
123
Soil moisture, which refers to soil humidity and water content in the soil, is the most vital requirement for plant growth, crop quality, cultivation, and maintenance of soil temperature [19, 20]. Soil moisture and temperature have to be adjusted and regulated precisely to avoid water overflow and the risk of evaporation and to protect the plants. Very low soil moisture withers the plants, adversely affects the photosynthesis, and in turn reduces the crop yield. On the other hand, extremely high soil moisture leads to blocking off crop breathing, weakening the plants roots, and increasing insect damage of crops. Insects directly damage the plants by eating leaves, fruit, and roots. Furthermore, insects affect crops indirectly by causing bacterial, viral, or fungal infection. The traditional irrigation cannot address these issues. As a result, proper water management and priority usage have to be optimally performed for agriculture to improve plant growth and crop yields, reduce manpower, and save water. In this paper, a Smart Heterogeneous Precision Agriculture (SHPA) system is proposed to measure the current agricultural parameters, predict subsequent agricultural parameter values, and select the sensing sensor nodes and sampling interval proactively. The SHPA system aims to filter measurement noise, increase sensing accuracy, improve energy efficiency, and enhance network lifetime. The paper is organized into eight sections including this one. After this introduction, Sect. 2 explores and critically analyses the related work into PA using WSNs. The problem definition and SHPA framework are formalized in Sects. 3 and 4, respectively. The extended Kalman filter, dynamic measurement, and energy models are introduced in Sect. 5. Then, Sect. 6 presents a detailed description of SHPA. A simulation-based evaluation is presented in Sect. 7. Finally, Sect. 8 summarizes the paper.
2 Related work Precision agriculture using WSNs is designed and developed in [21, 22]. In [21], a WSN based on IEEE 802.15.4 is developed to measure temperature. In [22], environmental parameters including temperature, air humidity, light, soil moisture, and pH level are monitored and controlled. Farmers can trigger the appropriate action based on the transmitted measurements. A Dynamic Bayesian Network (DBN) is employed to predict the environmental parameters. Agricultural monitoring systems using WSNs are developed in [23–26]. In [23], soil temperature, soil moisture, atmospheric temperature, atmospheric humidity, CO2 concentration, PH values, and illumination intensity are monitored. Alarms are sent when the parameters exceed threshold values. In [24], temperature, humidity, soil
Neural Computing and Applications
moisture, and pH of soil are measured for a potato field. Disease monitoring of grape vine crops is monitored in [25]. In [26], potato plants are monitored individually to improve crop production. Farmers can determine the requirements specifically for each potato plant, including the amount of fertilizer to be used and irrigation requirement. Automatic irrigation systems using WSNs have received the greatest attention. Soil moisture and temperature are measured and monitored in [1, 27, 28]. The system proposed in [29] measures temperature, solid moisture, humidity, and water tank level used for cardamom plant irrigation. In [30], soil moisture and temperature as well as air temperature and humidity, wind speed, and brightness are measured to reduce water usage and cost. Agricultural parameters, such as temperature, soil moisture and air humidity, are controlled and monitored in [31]. In [32], WSN is implemented as remote terminal unit to monitor and control water irrigation so that land productivity is increased. Soil moisture and temperature are measured and sent to control irrigation activity. Automated soil monitoring and sprinkler irrigation system is developed in [33] to achieve optimal water supply, yield, and crop health. Soil moisture and temperature, and pH are sensed and sent to the operator for monitoring purposes. An intelligent cultivation management system based on Zigbee WSN is presented in [34] to improve crop yield and quality and to get economic benefit. Temperature, humidity, light, and concentration of CO2 are sensed to monitor and control the garden and control crop growth conditions. In addition, the system sends alarm messages to the mobile phones of users through GSM network. In the above research about automatic irrigation systems, pumps/valves are opened automatically for irrigation when measurements exceed threshold values. Soil moisture and temperature are measured and monitored in [35]. An irrigation system is developed in [36] to monitor the crops and allow irrigation as needed. In [37], a model for a drip irrigation system is proposed to overcome the blockage of water emitters and broken pipes by monitoring soil moisture, temperature, and pressure. Soil water content for irrigation is measured in [38]. Threshold values for irrigation are predefined according to the crop type and plant growth stage. Water irrigation controller based on fuzzy logic is developed in [39] to compete the suitable time for irrigation based on the measured parameters sensed by the sensor node. Wireless Zigbee protocol is used to send the collected data locally to the farmer mobile phone, and GPRS link is adopted to send the real-time data to the remote station. Water management is considered in the literature also. The COMMON-Sense and iFarm systems are developed
for water management in [40, 41], respectively. In [42], the FLOW-AID system is designed under deficit conditions to monitor farm zones and send information to a decision support system in order to take appropriate actions. Researchers also consider PA within greenhouses. In [43], environmental conditions and growing process of cabbages and melons are monitored and controlled. In [44], environmental parameters like temperature and relative humidity are controlled. A Kalman Filter is used to estimate the states of the channel that are exploited to adjust the transmission power [45–47] consider measurement of temperature, humidity, and moisture. When the measurement threshold values are reached, pumps and valves are activated for irrigation. Temperature and mean humidity are measured and remotely transmitted using GSM for a large area in [48]. However, research in [48] addresses remote transmission of data from a large and far away area using GSM, rather than exploiting WSNs in the targeted area. In [49], a wireless mesh sensor network and novel routing algorithms are designed for sensing soil pH, electrical conductivity, and soil temperature and moisture. Nevertheless, [49] focuses on routing algorithms rather than the agricultural aspects. However, all of the above existing research about precession agriculture using WSNs adopts a fixed sampling interval for all crops. In [50], IEEE 802.15.4 MAC parameters are tuned using an analytical model so that the sampling interval is maximized to reduce the energy consumption of WSNs deployed for agriculture. On the other hand, the system used in [50] aims to maximize sampling interval without considering soil moisture and temperature, and water requirements of plants that vary according to plant age and type. Furthermore, the proposed technique in [50] is valid only for ZigBee WSNs to tune its MAC parameters. Most of the current research about PA considers the sensor measurement without the impact of noise associated with the measurements directly. In addition, sensor node selection for sensing is not considered. However, research in [51] considers the noise associated with measurements and selects the sensing sensor nodes so that the overall sensing accuracy of the whole agricultural land is optimized, subjected to the constraints of energy consumption and the maximum number of orphaned nodes. The system presented in [51] adopts random bit climbing (RBC) method to compute a near optimal sensor nodes selection. However, the network lifetime, energy efficiency, and proactive selection of the sampling interval are not addressed in [51]. Unlike previous research into precision agriculture that uses a fixed sampling interval equally for all crops, this research introduces a Smart Heterogeneous Precision Agriculture (SHPA) framework that provides additional
123
Neural Computing and Applications
advantages and benefits. SHPA aims to measure the current system state for online land monitoring and a decision support system. It also provides a smarter capability by predicting the next system state by which the system performance can be improved. Unlike previous research, the SHPA system considers the whole area of heterogeneous plants rather than considering a specific type of plant. Thus, it deals with each crop independently. The system also aims to minimize the estimation error of the system states to increase the certainty of actuators triggering at appropriate times by filtering the noise associated with measurements. Moreover, the proposed SHPA has the capability to protectively select the groups of sensor nodes that sense the agricultural parameters so that the network lifetime and sensing accuracy are maximized, subjected to predefine the sensor group size. Each plant is associated with a suitable sampling interval based on its agricultural requirements to increase the crop yields.
3 Problem definition As discussed in Sect. 2, the gathered sensing readings of agricultural parameters have a crucial role in decision support system process that is used to actuate the appropriate actions, such as irrigation, fertilizer dispensing, and insect combating. However, the sensing measurements suffer from associated noise that can cause the actuators to be triggered inappropriately. Consequently, the crops may be damaged and resources may be wasted. In addition, different sensor nodes may suffer from different amounts of noise according to their location with respect to the plant, each other, their residual energy, and their location within the network topology. Also, the relationship between the noise and the ‘‘pure’’ reading may be nonlinear. Therefore, SHPA considers the noisy measurements differently among the sensor nodes according to their state and condition. Furthermore, a precision agriculture system is considered as a dynamic system whose states vary and evolve with time according to nonlinear models. Agricultural parameters must be measured at the right times. A high sampling interval may lead to sensing the agricultural parameters after a long time and consequently result in damage to the crops. On the other hand, a small sampling interval may measure agricultural parameters unnecessarily and frequently leading to increased energy consumption and reduced network lifetime. Sensor node selection is another crucial aspect of WSNs. Sensor nodes should be selected to sense agricultural parameters so that reading accuracy and network lifetime are improved.
123
4 SHPA framework As shown in Fig. 1, an agriculture farm of width W and length L is divided into smaller areas in which the heterogeneous plants are located. A cluster-based topology is adopted in this research [52]. Each area is considered as a cluster where one cluster head (CH) is placed in the centre of the area. Other sensor nodes are scattered randomly across the farm. Sensors communicate directly with their cluster head for data transmission. The cluster heads relay on the data to the sink node that forwards it to the farmer’s server via Internet, cellular network, or satellite communication. The server sends the suitable commands to trigger actuators, such as valves and pumps for irrigation. Furthermore, the server performs the algorithms associated with SHPA. Since energy efficiency is a crucial feature in WSNs, the sensor node operates in one of predefined modes. These modes are as follows: sensing mode to sense soil moisture and temperature, communicating mode to transmit or receive data from other node, processing mode to perform processing tasks, and sleeping mode to listen for possible communication. The sensing, communicating, and processing modes are active modes where the sensor node consumes energy in sensing, transmitting or receiving, and processing activities respectively. On the other hand, the sleeping mode is a low-energy communication channel in which the sensor node listens to the channel status to initiate transmission or to wait for traffic from other sensor nodes. Therefore, the sensor nodes in the sleeping mode have to be triggered to ‘‘wakeup’’ if they are needed for communication, processing and/or sensing. Sensor nodes in sleeping mode use a low-energy communication channel [53] to receive a trigger message from other sensor nodes. The energy consumption to trigger the nodes to wakeup using the low-energy communication channel is neglected [53]. In this paper, the transpose of matrix A ¼ ½Arc will be denoted by AT ¼ ½Arc where Arc is the element at row ‘r’ and column ‘c’. The symbols i, j, and k identify the agricultural area, sensor node, and sensing time step, respectively. The farm is divided into ‘‘m’’ areas (ai ) where i ¼ 1; 2; . . .;m so that heterogeneous plants can be planted in the farm. Thus, the farm is a set of areas (ai ) as follows: A ¼ fa1 ; a2 ; . . .; am g. At time tai ðkÞ, each area (ai ) has a set Sai ðkÞ of ‘‘nai ðkÞ’’ sensor nodes as follows: Sai ðkÞ ¼ n o si1 ; si2 ; . . .; sinai ðkÞ where sij is the ‘‘jth’’ sensor node in area (ai ) and j ¼ 1; 2; . . .nai ðkÞ. Therefore, the set and total number of sensor nodes for the farm at time tai ðkÞ are defined as Snet ðkÞ and ‘‘NðkÞ’’, respectively, where Snet ðkÞ ¼ Sa1 ðkÞ [ Sa2 ðkÞ [ . . . [ Sam ðkÞ and
Neural Computing and Applications Fig. 1 SHPA framework
Farm Width
Farm Length
Internet, Cellular or Satellite
Servers
NðkÞ ¼
m P
Smart Phone
Laptop
Sensor
Plant
CH
Valves
Sink
Communication
nai ðkÞ. It is assumed that there are no conflicting
i¼1
nodes which are defined as sensor nodes that can sense agricultural parameters for more than one area. Therefore, Sap ðkÞ \ Saq ðkÞ ¼ U for all p m and q m where U is the empty set. The area and sensor node locations are defined T as Lai ¼ ½ xai yai T and Lsij ¼ xsij ysij , respectively. At each time step, the current agricultural parameters, predicted agricultural parameters, sensing sensor nodes, and sampling interval are determined.
5 The EKF, dynamic, sensor measurement and energy consumption models SHPA considers irrigation management by measuring soil moisture and temperature. At each time tai ðkÞ, soil moisture and temperature for each area (ai ) are measured and denoted as Mai ðkÞ and Tai ðkÞ, respectively. Therefore, the state vector at time tai ðkÞ of area (ai ) consists of the area soil moisture and temperature and is written as: X ai ðkÞ ¼ ½ Tai ðkÞ
Mai ðkÞ T
ð1Þ
5.1 Dynamic model The area ai dynamic model uses the discrete-time white noise acceleration model [54–56] which is described by: X ai ðk þ 1Þ ¼ f ½k; Xai ðkÞ þ wai ðkÞ
ð2Þ
where f ð:Þ is the evolution function or system transition function and is possibly a nonlinear and time-varying function that relates the current state X ai ðk þ 1Þ with the previous state X ai ðkÞ, and wai ðkÞ is the process noise which models the error in the system state values and is assumed
to be a zero-mean white Gaussian distribution process with Qai ðkÞ covariance matrix such that wai ðkÞ N 0; Qai ðkÞ .
5.2 Sensor measurement model The measurement model that relates the noisy measurements with the system state at time tai ðk þ 1Þ is given by: Zai ðk þ 1Þ ¼ hai ½k þ 1; X ai ðk þ 1Þ þ vai ðk þ 1Þ
ð3Þ
where hð:Þ is the measurement function and is possibly a nonlinear and time-varying function and vai ðk þ 1Þ is the measurement noise whose known distribution is independent from process noise and time, and is assumed to possess a zero-mean white Gaussian distribution with Rai ðk þ 1Þ covariance matrix such that vai ðk þ 1Þ Nð0; Rai ðk þ 1ÞÞ.
5.3 Extended Kalman filter (EKF) The extended Kalman Filter (EKF) [54–56] is based on the linearization of the nonlinearities in dynamic and/or measurement models. EKF is a mathematical model that can use the noisy measurements to estimate values that are close to the true values. It deals with conditional mean and covariance. EKF is used to calculate the predicted and updated states and their covariance matrices. In the prediction stage, the predicted state vector (X^ai ðk þ 1jkÞ) with its associated predicted covariance matrix (Pai ðk þ 1jkÞ) is calculated at time tai ðkÞ. In the update stage, the update state vector X^ai ðk þ 1jk þ 1Þ with its associated updated covariance matrix Pai ðk þ 1jk þ 1Þ is calculated at time tai ðk þ 1Þ. In the prediction stage at time tai ðkÞ, the predicted state vector is given by: X^ai ðk þ 1jkÞ ¼ f k; X^ai ðkjkÞ ð4Þ
123
Neural Computing and Applications
with associated predicted covariance matrix given by: Pai ðk þ 1jkÞ ¼ Fai ðkÞPai ðkjkÞFTai þ Qai ðkÞ
where Fai ðkÞ is the Jacobian matrix of f ½k; X ai ðkÞ at X ai ðkÞ ¼ X^ai ðk þ 1jkÞ as follows: Fai ðkÞ ¼
of ½k; X ai ðkÞ ; at X ai ðkÞ ¼ X^ai ðk þ 1jkÞ oX ai ðkÞ
ð7Þ
with associated residual or innovation covariance matrix: Sai ðk þ 1Þ ¼ Rai ðk þ 1Þ þ Hai ðk þ 1ÞPai ðk þ 1jkÞHTai ðk þ 1Þ ð8Þ where ð9Þ
the Jacobian matrix X ai ðk þ 1Þ ¼ X^ai ðk þ 1jkÞ
oh½k þ 1; X ai ðk þ 1Þ ; at Xai ðk þ 1Þ oX ai ðk þ 1Þ ¼ X^ai ðk þ 1jkÞ
of as
Hai ðk þ 1Þ ¼
TX v
Electronics
v
L bits
v
Electronics d
2 2*L*d
3*L
Fig. 2 Energy consumption model
wTX ðl; dij Þ ¼ a1 l þ a2 l dij2
ð14Þ
where a1 is the electronic energy required to transmit one bit that depends on factors, such as coding, modulating, and filtering, and a2 is related to the radio energy (i.e., the amplifier energy required to transmit one bit in free space). The power loss in free space is proportional inversely to square of the distance. Therefore, it is modelled in Eq. (14) by multiplying with dij2 . Thus, the units for a1 and a2 are Joule/bit (J/b) and Joule/bit/meter2 (J/b/m2). The energy consumption to receive an l-bit message is: wRX ¼ a3 l
ð15Þ
where a3 is the amplifier energy required to receive one bit from free space. The unit for a3 is Joule/bit (J/b).
6.1 SHPA monitoring algorithm
ð11Þ with associated updated covariance matrix: ð12Þ
where the filter gain is given by:
At each time tai ðkÞ, SHPA aims to select a sensing set (Ss ðai ; kÞ) for each area of ‘‘ns ðai ; kÞ’’ sensor nodes to measure the system state, where Ss ðai ; kÞ ¼ si1 ; si2 ; . . .; sins ðai ;kÞ , ns ðai ; kÞ nai ðkÞ and Ss ðai ; kÞSai ðkÞ. Therefore, the state noise measurement vector Zai ðk þ 1Þ and measurement function hai ½k þ 1; X ai ðk þ 1Þ of Eq. (3) are given by: h i Zai ðk þ 1Þ ¼ zsi1 ðk þ 1Þ zsi2 ðk þ 1Þ . . . zsins ðai ;kþ1Þ ðk þ 1Þ
ð13Þ
ð16Þ 2
5.4 Energy consumption model Energy is consumed during sensing, communicating, and processing activities. As shown in Fig. 2, the transmitter consumes energy to run the radio electronics and the power amplifier, while the receiver consumes energy to run the radio electronics [57, 58]. The energy consumption of sensor si to transmit l-bit message to the sensor sj over a distance dij is:
123
RX
Amplifier
ð10Þ
X^ai ðk þ 1jk þ 1Þ ¼ X^ai ðk þ 1jkÞ þ K ai ðk þ 1ÞCai ðk þ 1Þ
K ai ðk þ 1Þ ¼ Pai ðk þ 1jkÞHTai ðk þ 1ÞS1 ai ðk þ 1Þ
v
Receiver (RX)
6 Detailed description of SHPA system
The update state estimate is given by:
Pai ðk þ 1jk þ 1Þ ¼ Pai ðk þ 1jkÞ K ai ðk þ 1ÞSai ðk þ 1ÞK Tai ðk þ 1Þ
TX
1*L
Cai ðk þ 1Þ ¼ Zai ðk þ 1Þ Z^ai ðk þ 1jkÞ
and Hai ðk þ 1Þ is h½k þ 1; X ai ðk þ 1Þ at follows:
L bits
ð6Þ
In the update stage at time tai ðk þ 1Þ, where the measurements at time tai ðk þ 1Þ are available, the measurement residual, which is the difference between actual and predicted measurements, is calculated as follows:
Z^ai ðk þ 1jkÞ ¼ hai k þ 1; X^ai ðk þ 1jkÞ
Transmitter (TX)
ð5Þ
6 6 hai ½k þ 1; X ai ðk þ 1Þ ¼ 6 6 4
hai ½k þ 1; X ai ðk þ 1Þ; si1 hai ½k þ 1; X ai ðk þ 1Þ; si2 .. . hai k þ 1; X ai ðk þ 1Þ; sins ðai ;kþ1Þ
3 7 7 7 ¼ ½hi1 7 5
ð17Þ for 1 i ns ðai ; kÞ where zsij ðk þ 1Þ ¼ Tsij ðk þ 1Þ
Msij ðk þ 1Þ
T
ð18Þ
Neural Computing and Applications
Thus, the Jacobian matrix of h½k þ 1; Xai ðk þ 1Þ at X ai ðk þ 1Þ ¼ X^ai ðk þ 1jkÞ based on Eq. (10) is determined by: Hai ðk þ 1Þ ¼ Hij 1 i 2ns ðai ; kÞ; 1 j 2 ð19Þ where Hij ¼
8 ohi1 ½k þ 1; X ai ðk þ 1Þ > > < oT ðk þ 1Þ
at
X ai ðk þ 1Þ ¼ X^ai ðk þ 1jkÞ
j¼1
> oh ½k þ 1; X ai ðk þ 1Þ > : i1 oMai ðk þ 1Þ
at
X ai ðk þ 1Þ ¼ X^ai ðk þ 1jkÞ
j¼2
ai
ð20Þ Similarly, the Jacobian matrix Fai ðkÞ of f ½k; X ai ðkÞ at Xai ðkÞ ¼ X^ai ðk þ 1jkÞ in Eq. (6) is determined as: Fai ðkÞ ¼ Fij 1 i 2; 1 j 2 ð21Þ where 8 ofi1 > > < oTai ðkÞ Fij ¼ ofi1 > > : oMai ðkÞ
planted in the area centre. The contribution of the sensor node to the network lifetime LT k; ai ; sij is defined as the sum of its remaining energy and the number of its neighbouring nodes. For each sensor node (sij ) in the area (ai ), the contributions of sensing accuracy and the network lifetime are modelled as follows: 1 EID sij ; ai ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð23Þ T Lsij Lai Lsij Lai 2 3 6 Er ðk; ai ; sij Þ þ LTðk; ai ; sij Þ ¼ 4 P Er ðk; ai ; sil Þ 8l2Sai ðkÞ
Gðk; ai ; sij Þ 7 P 5 Gðk; ai ; sil Þ 8l2Sai ðkÞ
at
X ai ðkÞ ¼ X^ai ðk þ 1jkÞ j ¼ 1
ð24Þ where Er k; ai ; sij is the remaining energy of sensor node sij at time step k and G k; ai ; sij is the number neighbouring nodes of sensor node sij at time step k which is calculated as follows:
at
Xai ðkÞ ¼ X^ai ðk þ 1jkÞ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
T G k; ai ; sij ¼ countsil 2Sai ðkÞ Lsij Lsil Lsij Lsil Rr
j¼2 ð22Þ
The SHPA system monitors the agricultural land by measuring the system state that consists of soil moisture and temperature. It comprises two stages: update and prediction stages. In the update stage at time tai ðkÞ, the selected sensing sensor set sense the measurements. Then, the updated and predicted system states with its associated matrices are calculated according to the available measurements up to time tai ðkÞ.
6.2 Sensing node selection algorithm (SNSA) for SHPA For each time step tai ðk 1Þ and for each area (ai ), SHPA invokes the SNSA to proactively select the set of sensing sensor nodes, Ss ðai ; kÞ that will sense the agricultural parameters at the next time step tai ðkÞ. SNSA aims to maximize sensing accuracy and network lifetime. Therefore, the objective function for each sensor node consists of the weighted sum of the contributions in sensing accuracy and network lifetime. The sensing sensor nodes selection objective function is maximized, subject to predefine constraints. Thus, the sensing accuracy of the sensor node EID ai ; sij is defined as the inverse of the Euclidean distance between the sensor node and the crop which is
for 8l 6¼ j
ð25Þ Therefore, at time step k, the objective function is defined as: Fobj ½k; ai ; sij ¼ bðk; ai Þ
EIDðsij ; ai Þ P EIDðsil ; ai Þ 8l2Sai ðkÞ
þ ½1 bðk; ai Þ LTðk; ai ; sij Þ
ð26Þ
The ‘th sensor node, g‘ 2 Ss ðai ; kÞ where ‘ ns ðai ; kÞ is selected as follows ð27Þ g‘ ¼ argsij max Fobj k; ai ; sij : sij 2 Sai ðkÞngr Subject to: ð1Þ ns ðai ; kÞ ¼ nc
for all i
ð2Þ ns ðai ; kÞ nai ðkÞ for all i
ð28aÞ ð28bÞ
where nc is a predefined group size and Sai ðkÞngr is the set of Sai ðkÞ members excluding the ones from g1 to g‘1 . Algorithm 1 shows the sensing node selection algorithm (SNSA) running at each area.
123
Neural Computing and Applications
7 Simulation results
6.3 Complete SHPA algorithms Algorithm 2 shows the complete SHPA algorithm. It starts with initialization routine to set the initial values of time step, states, covariance matrices, and sampling interval. After that, the EKF starts with the update stage to calculate the updated states and is associated covariance matrix. Then, the prediction stage is performed to predict the predicted states and is associated covariance matrix. SNSA is achieved in the prediction stage to proactively select the sensor node for the next update stage.
123
In this section, the performance of the proposed SHPA system is evaluated using event-driven simulation. C?? is used to build the simulation environment using an Intel Core i5 2.5 GHz processor and 4 GB memory. The simulation results are averaged over 20 runs with different random sensor placement using a fixed sensor density, and the simulation duration is 3 days. The farm of 200 m 9 200 m is divided into three equal areas (m ¼ 3) of 100 m 9 100 m for each area. Four cluster heads are placed in the centre of each area, and one sink node is placed in the most left-top corner of the farm. Unless specifically mentioned, for each time step, four sensor nodes (nc ¼ 3) are selected for each area to sense soil moisture and temperature. The sensor reading is assumed to be within a sensing range of 50 m (Rr). Therefore, to increase likelihood of sensing the area by three sensor nodes with sensing range of 50 m at all-time steps, the number of sensor nodes deployed in the agricultural farm should be calculated properly. Three sensor nodes or more within the sensing range can be guaranteed by using sensor density of q ¼ 5:6 103 sensors=m2 [59]. Thus, for a given agricultural farm of A ¼ 200 200 m2 , the number of required sensor nodes to be uniformly deployed is N ¼ qA ¼ 224 sensor nodes. Unless specifically mentioned, the weighting parameter, b of SNSA in Eq. (26) is set to 0.5. We assume line of sight (LOS) communication between the nodes within the same coverage area. Two nodes are in the same coverage area if the distance between them is equal to or less than the radio range, which is set to 100 m. Therefore, the cluster head for each area is within the coverage area of all sensor nodes in the same area. The cluster head is also within the coverage area of its neighbouring cluster heads. The destination-sequenced distance vector routing (DSDV) routing protocol, which is a proactive ad hoc routing protocol, has been implemented in the simulation model [60]. It is based on classical distributed Bellman–Ford (DBF) algorithm. In DBF, each sensor node maintains the first sensor node (hop) on the shortest path to every other sensor node in the network. Each sensor node maintains routing table for all possible destinations and the number of routing hops to reach that destination. A sequence numbering system (labelling the routes) is used to differentiate stale routes from the new routes. The energy model parameters are set as follows: a1 = 100 nJ/b, a2 = 1 pJ/b/m2, a3 = 100 nJ/b, and the sensing energy cost for all sensor nodes is assumed to be 8 nJ. The energy level of sensor nodes is set randomly to a value between 3 and 13 J. The communication load for all measurements and management messages is assumed to be l ¼ 200 KB. For the simulation purposes [51], the system
Neural Computing and Applications
is assumed to evolve according to the following evolution function: 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 3 0:8Tai ðkÞ þ 0:8pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi jTai ðkÞ 80ffij f 6 7 f ½k; Xai ðkÞ ¼ 11 ¼ 4 jMai ðkÞ 80j 5 f21 0:96Mai ðkÞ þ Dtai ð29Þ with covariance matrix of the process noise given 0:5 0 by:Qai ðkÞ ¼ for all areas, ai at all times. 0 0:2 Therefore, the Jacobian matrix Fai ðkÞ of f ½k; Xai ðkÞ at X ai ðkÞ ¼ X^ai ðk þ 1jkÞ based on Eqs. (21) and (22) is equal to: F11 F12 Fai ðkÞ ¼ : ð30Þ F21 F22 where F11 ¼
of11 0:4 ¼ 0:8 þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi oTai ðkÞ jTai ðk þ 1jkÞ 80j
0:5 , F22 ¼ oMofai21ðkÞ ¼ 0:96 þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi , F12 ¼ oMofai11ðkÞ ¼ 0, and jMai ðkÞ80j F21 ¼ oTofai21ðkÞ ¼ 0: The measurement function for each sen-
sor node, sij , of the measurement model is assumed to be [51]: hai k þ 1; Xai ðk þ 1Þ; sij ¼ ½ h11 h21 T pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi T ¼ Mai ðk þ 1Þ Ta i ð k þ 1 Þ ð31Þ where measurement noise variance of the sensor node, selected to sense soil moisture and temperate, is set to between a value of 10 to 15 based on the sensor distance from the area centre or crop root. Therefore, the Jacobian matrix Hai ðk þ 1Þ of h½k þ 1; Xai ðk þ 1Þ at X ai ðk þ 1Þ ¼ X^ai ðk þ 1jkÞ based on Eqs. (19) and (20) is equal to: H11 ¼ H31 ¼ Hð2ns ðai ;kÞ1Þ;1 ¼
ohi1 oTai ðk þ 1Þ
1 ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 Tai ðk þ 1jkÞ for i ¼ 1; 3; . . .; ð2ns ðai ; kÞ 1Þ; ohi1 1 ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi oMai ðk þ 1Þ 2 Mai ðk þ 1jkÞ for i ¼ 2; 4; . . .; 2ns ðai ; kÞ;
H22 ¼ H42 ¼ H2ns ðai ;kÞ;2 ¼
ohi1 ¼0 oTai ðk þ 1Þ for i ¼ 2; 4; . . .; 2ns ðai ; kÞ; and
H21 ¼ H41 ¼ H2ns ðai ;kÞ;1 ¼
ohi1 ¼0 oMai ðk þ 1Þ for i ¼ 1; 3; . . .; ð2ns ðai ; kÞ 1Þ
H12 ¼ H32 ¼ Hð2ns ðai ;kÞ1Þ;2 ¼
In this section, the following performance metrics are used to evaluate the proposed SHPA. The energy balance of area (EBðai Þ) is defined as the mean of normalized energy remaining for sensor nodes inside the area at the end of simulation time (Ke ) which is calculated as follows: Pnai ðKe Þ r ½E ðKe ; ai ; sil Þ=Em ð0; ai ; sil Þ ð32Þ EBðai Þ ¼ l¼1 nai ðKe Þ where Er ðKe ; ai ; sil Þ is the remaining energy of sensor, sil at the end of simulation time (Ke ), and Em ð0; ai ; sil Þ is the maximum energy of sensor (sil ) at the initial time. The neighbour nodes count of area (NCðai Þ) is defined as the average of neighbour nodes count for sensor nodes inside the area at the end of simulation time (Ke ) which is calculated as follows: P n ai ð Ke Þ GðKe ; ai ; sil Þ ð33Þ NCðai Þ ¼ l¼1 nai ðKe Þ where nai ðKe Þ is the sensor nodes number of area (ai ) at the end of simulation time (Ke ). The neighbour nodes count, energy balance, and dead nodes count are used as indicators for network lifetime. Small value of neighbour nodes count means that the neighbours of sensor nodes are small, and hence, the prevalence of gaps in the network is increased, and consequently, the network lifetime is reduced. The normalized energy remaining of the nodes that are not selected for sensing is a unity as there is no energy dissipated from these nodes. Therefore, high value of energy balance indicates that less nodes are engaged for operation, and hence, the energy load balancing among senor nodes is degraded which causes a quick death for senor nodes and reduces network lifetime. As a result, improving the network lifetime requires high values of neighbour nodes count and energy balance, and low values of dead nodes count.
7.1 SHPA evaluation Figure 3 shows the evaluation of SHPA approach. The agricultural farm with four areas is shown in Fig. 3a. The sensor node and cluster head locations with routing paths based on DSDV are also shown. Each cluster head selects the shortest path through other cluster heads to send messages to sink node. For example, the path for the CH (2) to send message to sink node is [CH (2)–CH (3)–Sink Node]. The heterogeneous sampling interval versus the simulation time for each area is plotted in Fig. 4. According to
123
Neural Computing and Applications 30
Sampling Interval (min)
25
20
15
10 Area (1) Area (2) Area (3) Area (4)
5
0
0
500
1000
1500
2000
2500
3000
3500
4000
3500
4000
Simulation Time (min)
Fig. 4 Sampling interval for each area 30
Atmosphere Temperature ( °C )
the SHPA algorithm shown in Algorithm 2, the sampling interval is set independently for each area according to its requirements. Thus, Fig. 4 shows that the sampling interval for the area (1), (2), (3) and (4) is set to the values 10, 15, 20, and 25 min, respectively, during the simulation time. At each time step (k), soil moisture (Mai ðkÞ) for each area (ai ) and atmospheric temperature (Tai ðkÞ) for agricultural farm are measured and plotted in Figs. 5 and 6, respectively. Atmospheric temperature and soil moisture follow the dynamic model assumed in Eq. (29). As shown in Fig. 6, the higher of sampling interval is, the more reduction of soil moisture of the area occurs because the water evaporation of soil increases with the increasing of sampling interval. Therefore, the proposed SHPA approach considers water requirements for each crop individually so that the agricultural production is maximized. The updated and predicted covariance errors for each area with the simulation time shown in Fig. 7 are defined as the ‘‘trace’’ of the updated covariance matrix Pai ðk þ 1jk þ 1Þ and the predicted covariance matrix, Pai ðk þ 1jkÞ, respectively. As shown in Fig. 7, the updated and predicted covariance errors are initially high and are reduced with more gathered measurements. Figure 8 shows the total energy consumption versus the simulation time for each area. For all areas, the energy consumption increases with the simulation time because each sensing activity consumes energy in sensing and communicating. As shown in Fig. 8, with decreasing sampling interval, the increase in energy consumption with simulation time increases. Therefore, area (1) and area (4) have the most and the lowest energy consumption as they have the lowest and the highest sampling intervals, respectively.
29.5
29
28.5
28
27.5
27
0
500
1000
1500
2000
2500
3000
Simulation Time (min)
Fig. 5 Atmospheric temperature for land 60
Sink 200 180
50
160
CH(3)
CH(4)
Soil Moisture (%)
Farm Width (m)
Area (1) Area (2) Area (3) Area (4)
140 120 100 80
40
30
20
60 40
CH(1)
CH(2)
10
20 0
0 0
20
40
60
80
100
120
Farm Length (m )
Fig. 3 Network topology and structure
123
140
160
180
200
0
500
1000
1500
2000
2500
3000
Simulation Time (min)
Fig. 6 Soil moisture of SHPA for each area
3500
4000
Neural Computing and Applications 600
In this section, the robustness of SHPA is evaluated through considering different number of areas (m). In Fig. 9, total energy consumption, neighbour nodes count, dead nodes count, and predicted and updated covariance errors are plotted with areas count of 1, 2, 4, 6, 4, and 10. With increasing the number of areas, more sensor nodes are involved to sense agricultural parameters and to send the measurements. Hence, the total energy consumption and dead nodes count increase with increasing the number of areas. In Fig. 10, the total energy consumed during the simulation time is plotted for different number of areas. With increasing the dead nodes count, the neighbour nodes count reduces, as shown in Fig. 9. Nevertheless, when the number of areas increases from 1 to 10 (i.e., 90% from its maximum value), neighbour nodes count reduces from 10.6 to 8.5 (i.e., 19.8% from its maximum value). This indicates that SHPA can maintain its performance in terms of neighbour nodes count and network lifetime for different number of areas. The predicted and updated covariance errors are not significantly changed with increasing the number of areas.
500
Energy Consumption (J)
7.2 The impact of number of areas
300
200
0
0
500
Covariance Error
1
1000
2000
3000
2
Covariance Error
Covariance Error
1
1000
4000
0
2000
3000
Simulation Time (min)
1000
2000
3000
4000
Simulation Time (min)
Area (3)
0
3500
1
0
4000
3
0
3000
Area (2)
Simulation Time (min)
2
2500
3
Area (1)
0
2000
section. The selected group size is set to fixed values of 1, 2, 3, 4 and 5. In Fig. 11, total energy consumption, neighbour nodes count, dead nodes count, and predicted and updated covariance errors are plotted with the selected group size. With increasing the group size, more sensor nodes are contributed to measure the agricultural parameters and send and receive the measurement readings. Consequently, the energy consumption rises with increasing the group size because the sensing and communication activities increase with the selected group size. In Fig. 12, the total energy consumed during the simulation period is
Predicted 2
1500
Fig. 8 Energy consumption for each area
Updated
3
0
1000
Simulation Time (min)
The effect of the selected group size, nc , defined in Eq. (28) to read the moisture and temperature is evaluated in this
Covariance Error
400
100
7.3 The impact of selected group size
Fig. 7 Updated and predicted covariance error for each area
Area (1) Area (2) Area (3) Area (4)
4000
3
2
Area (4)
1
0
0
1000
2000
3000
4000
Simulation Time (min)
123
Neural Computing and Applications 5000
12
Neighbour Count
Energy Consumption (J)
Fig. 9 Impact of areas count in SHPA performance
4000 3000 2000 1000 2
4
6
8
10
8
6
10
2
Number of Areas
4
6
8
10
Number of Areas
Covariance Error
30
Dead Count
25 20 15 10
1.2 Updated
1
Predicted 0.8 0.6
5
2
4
6
Nmber of Areas
Number of Areas=1 Number of Areas=2 Number of Areas=3 Number of Areas=4 Number of Areas=5 Number of Areas=6
Energy Consumption (J)
3500 3000 2500 2000 1500 1000 500 0
0
500
1000
1500
2000
2500
3000
3500
4000
Simulation Time (min)
Fig. 10 Total energy consumption of the agricultural farm for different areas count
plotted for different group size. Neighbour nodes and dead nodes counts shown in Fig. 11 are degraded with group size because with increasing the group size, more sensor nodes are involved in sensing activities and in turn more sensor nodes dies. On the other hand, the predicted and updated covariance errors are improved with increasing the group size because more measurements are collected with increasing the group size, and in turn the EKF can filter the noises more efficiently.
123
10
2
4
6
8
10
Nmber of Areas
7.4 Comparison with other approaches
4500 4000
8
The proposed SHPA approach is compared with most current research approaches that gather the sensor reading without using EKF. Figures 13 and 14 show moisture and temperature errors versus the simulation time for different areas with and without SHPA. In case of ‘‘with SHPA’’, moisture and temperature errors are defined as the difference between the real system state and EKF updated states. In case of ‘‘without SHPA’’, moisture and temperature errors are defined as the difference between the real state and sensor noisy measurements. As shown in Figs. 13 and 14, moisture and temperature errors are lower in case of ‘‘with SHPA’’ approach, compared with ‘‘without SHPA’’ approach because the EKF can filter the noise from the noisy readings effectively while the random noise is associated with sensors readings in case of ‘‘without SHPA’’ approach. The RBC method is introduced in [51] to select the sensor nodes that nearly maximizing the sensing accuracy. The noise associated with the measurements is filtered using EKM. Dynamic and measurement models for soil temperature and soil moisture are considered in [51]. As shown in Sect. 6.2, the proposed SHPA introduces novel objective function and heuristic approach to select the sensor nodes to measure soil temperature and soil moisture so that sensing accuracy and network lifetime are improved. SHPA is adopted EKF to eliminate the noise associated with the measurements and to predict the next
Neural Computing and Applications 9.8
2000
Neighbour Count
Energy Consumption (J)
Fig. 11 Impact of selected group size in SHPA performance of the agricultural farm
1500
1000
500 1
2
3
4
9.6 9.4 9.2 9 8.8 8.6
5
1
2
Group Size
Covariance Error
Dead Count
5 4 3
1
2
3
4
Group Size
Group size=1 Group size=2 Group size=3 Group size=4 Group size=5
Energy Consumption (J)
1600 1400 1200 1000 800 600 400 200 0
0
500
1000
1500
2000
2500
3000
5
2
6
1800
4
Group Size
7
2
3
3500
4000
Simulation Time (min)
Fig. 12 Total energy consumption of the agricultural farm for different group sizes
system states. Therefore, the proposed SHPA is compared with RBC method introduced in [51]. The performance metrics averaged over the four areas is shown in Table 1 for the proposed SHPA and the RBC schemes [51]. The sampling interval for all areas is set to 10 min. In Table 1, the performance metrics of SHPA are calculated for different weighting parameter (b) of 0, 0.5, and 1. According to Eq. (26), increasing weighting parameter (b) leads to increase the chance to select nodes with small associated noise and, in turn, the RMSE of T and M is improved with increasing b. On the other hand,
5
Updated Predicted
1.5
1
0.5
1
2
3
4
5
Group Size
the performance of neighbour nodes count, energy balance, and dead nodes count degrades with increasing b. The objective of the selection method in [51] only aims to minimize estimation error. RBC is adopted in [51] to get the near optimal solution. Nevertheless, as introduced in Sect. 6.2, the proposed SNSA aims to select the sensor nodes so that network lifetime and sensing accuracy are maximized. Thus, as shown in Table 1, compared with RBC scheme, the proposed SHPA scheme at b ¼ 0 improve the performance metrics of neighbour nodes count, energy balance, dead nodes count, and consumed energy. However, at b ¼ 0, the RMSE of T and M is silently degraded, compared with SHPA. The RMSE of T and M can be improved in SHPA by setting b to unity. By setting b to 0.5, the SHPA scheme can still perform better than the RBC scheme, where the performance metrics of neighbour nodes count, energy balance, dead nodes count, and consumed energy are improved. Therefore, the weighting parameter (b) is a design and controlled parameter which controls the weight of improving RMSE and network lifetime. The weighting parameter (b) can be selected in-line during the system running based on the current system states and requirements.
7.5 Computational complexity analysis Recalling Sect. 6.2, the proposed SHPA aims to select nc sensor nodes for each area by using quicksort algorithm. Therefore, the computational complexity of the proposed SHPA algorithm is Oðn log nÞ. On the other hand, the RBC
123
Neural Computing and Applications Fig. 13 Moisture error of each area with and without SHPA
8
Moisture Error (%)
Moisture Error (%)
8 6 4 2 0
Area (1) 0
1000
2000
3000
6 4 2 0
4000
Area (2) 0
Simulation Time (min)
1000
2000
3000
4000
Simulation Time (min) with SHPA
6 4 2 0
without SHPA
8
Moisture Error (%)
Moisture Error (%)
8
Area (3) 0
1000
2000
3000
6 4 2 0
4000
Area (4) 0
Simulation Time (min)
4000
Temperature ( °C )
4
2
Area (1)
0
1000
2000
3000
6
4
2
0
4000
Area (2)
0
Simulation Time (min)
1000
2000
3000
4000
Simulation Time (min) 8
Temperature ( °C )
8
Temperature ( °C )
3000
8
6
0
2000
Simulation Time (min)
8
Temperature ( °C )
Fig. 14 Temperature error of each area with and without SHPA
1000
6
4
2
Area (3)
6
4 with SHPA 2
without SHPA Area (4)
0
0 0
1000
2000
3000
Simulation Time (min)
approach selects nc sensor nodes for each area. For each chromosome, the updated covariance matrix error is calculated for each area separately and then it sums up to get the objective function. The process is repeated based on the number of allowed iterations. Hence, the computational
123
4000
0
1000
2000
3000
4000
Simulation Time (min)
complexity of RBC scheme is Oðn3 Þ. Therefore, it is clear that RBC is computationally more complex than the proposed SHPA.
Neural Computing and Applications Table 1 Performance metrics of agricultural farm using SHPA and RBC schemes Scheme
Neighbour nodes count
Energy balance
Dead nodes count
Consumed energy (J)
RMSE of T (°C)
RMSE of M (%)
SHPA b=0
10.09
0.57
0.10
435.7
0.138
0.177
b = 0.5
9.11
0.62
5.35
435.1
0.133
0.174
b=1
6.62
0.94
25.81
434.4
0.132
0.171
7.89
0.63
14.09
435.8
0.134
0.172
RBC
7.6 Discussion
8 Conclusion
As shown from the results, the proposed SHPA approach provides several benefits and advantages. As shown in Figs. 5 and 6, SHPA achieves a remote monitor and control system for precision agriculture. Figures 13 and 14 show that SHPA can be able to filter the noise associated with the noisy measurements by adopting EKF. A group of sensor nodes is selected to measure the agricultural parameters with the aim of increasing the sensing accuracy which is shown in Table 1. Furthermore, SHPA can predict the next state of the system using the history of measurements and the system dynamic model. Hence, the agricultural activities, such as water irrigation, are performed at precise time and location with accurate amount. Moreover, crop production and water saving are improved. As clear from Table 1, SHPA approach considers also the improvement of network lifetime in the sensor node selection as another main objective beside sensing accuracy enhancement. Since a certain number of sensor nodes are chosen to measure agricultural parameters, redundant readings and data transmission are minimized, and, in turn, the total energy consumption is reduced. Therefore, the time consumed to discharge the battery of a sensor node is increased which causes an improvement of sensor node lifetime. Heterogeneity is another main feature of the proposed SHPA scheme. SHPA supports different crop and soil types by selecting different sampling interval for each crop as shown in Fig. 4. In addition, by adjusting the weighting parameter of objective function mention in Eq. (26), SHPA can cope with different operational conditions. The complexity of the proposed SHPA increases with the size of sensor nodes group. Nevertheless, as explained in Sect. 4, the SHPA performs its algorithms using remote server with high resource capabilities. In addition, the sensor node is assumed to measure the agricultural parameters for a certain area without considering sensor nodes that can give readings from the borders of different areas. However, this is realistic since the farm is assumed to be divided into areas with different crop and soil types.
This paper explores and critically analyses the current research about the precision agriculture using WSNs. SHPA approach is introduced in this paper to overcome the limitations and enhance approaches of other research. The agricultural farm is divided into smaller areas in which different type of plants are implanted. SHPA employs the EKF to precisely calculate the current agricultural parameters including soil moisture and temperature and predicts the next ones so that crop yields are improved. System dynamic and measurement model is presented for the purposes of this research. Also, SNSA is introduced to select the sensing sensor nodes efficiently and proactively at each time step so that network lifetime and sensing accuracy are improved. In this research, heterogeneous sampling interval is introduced so that each crop is associated with the appropriate sampling interval according to its agricultural requirements, where the crop yields are maximized. The simulation results show that the proposed SHPA scheme eliminates the noise associated with the measurements, compared with other schemes. Therefore, the agricultural activities, such as irrigation, can be performed in a precise time to enhance the crop yield. In addition, SHPA scheme improves the performance in terms of neighbour nodes count, energy balancing, dead nodes count, and consumed energy, compared with other schemes. Consequently, network lifetime and sensing accuracy for the introduced SHPA are enhanced. Furthermore, it is shown that the proposed SHPA system is less complex computationally, compared with RBC scheme. Acknowledgements We would like to express our thankfulness to Ibhath Project (Qatar Charity at Palestine) which funds this research.
Compliance with ethical standards Conflict of interest The authors declare that they have no conflict of interest.
123
Neural Computing and Applications
References 1. Tarange PH, Mevekari RG, Shinde PA (2015) Web based automatic irrigation system using wireless sensor network and embedded Linux board. In: IEEE international conference on circuit, power and computing technologies (ICCPCT), India, pp 1–5 2. Panchard J, Prabhakar TV, Hubaux J, Jamadagni HS (2007) Common sense net: a wireless sensor network for resource-poor agriculture in the semi-arid areas of developing countries. Inf Technol Int Dev 4(1):51–67 3. Kumar SA, Ilango P (2017) The impact of wireless sensor network in the field of precision agriculture: a review. Wireless personal communications. Springer, Berlin, pp 1–14 4. Wieselthier JE, Nguyen GD, Ephremides A (2002) Resource management in energy-limited, bandwidth-limited, transceiverlimited wireless networks for session based multicasting. Comput Netw: Int J Comput Telecommun 39(5):113–131 5. Potdar V, Sharif A, Chang E (2009) Wireless sensor networks: a survey. In: IEEE international conference on advanced information networking and applications workshops, UK, pp 636–641 6. Zhao F, Liu J, Liu J et al (2003) Collaborative signal and information processing: an information directed approach. Proc IEEE 91(8):1199–1209 7. Akyildiz IF, Su W, Sankarasubramaniam Y et al (2002) Wireless sensor networks: a survey. Comput Netw: Int J Comput Telecommun Networ, ACM 38(4):393–422 8. Estrin D, Govindan R, Heidemann J et al (1999) Next century challenges: scalable coordination in sensor networks. Proc MobiCom, ACM, pp 263–270 9. Arampatzis T, Lygeros J, Manesis S (2005) A survey of applications of wireless sensors and wireless sensor networks. In: Proceedings of the IEEE 13th mediterranean conference on control and automation, cyprus, pp 719–724 10. Crary SB, Baer WG, Cowles JC et al (1990) Digital compensation of high-performance silicon pressure transducers. Sens Actuators A 21(2–3):70–72 11. Paek J, Hicks J, Coe S, Govindan R (2014) Image-based environmental monitoring sensor application using an embedded wireless sensor network. Sens (Basel) 14(9):15981–16002 12. Leea WS, Alchanatis V, Yang C et al (2010) Sensing technologies for precision specialty crop production. Comput Electron Agric 74(1):2–33 13. Adam ML, Cook S, Corner R (2000) Managing uncertainty in site-specific management: what is the best model? Precision Agric 2(1):39–54 14. Yazdi N, Mason A, Najafi K et al (2000) A generic interface chip for capacitive sensors in low-power multi-parameter microsystems. Sens Actuators, A 84(3):351–361 15. Lee K (1998) A smart transducer interface for sensors and actuators. Smart Transducer Interface Standard-IEEE 1451.2 Standard 16. Leon CT, Shaw DR, Cox MS et al (2003) Utility of remote sensing in predicting crop and soil characteristics. Precis Agric. Kluwer Acad Publ 4(4):359–384 17. Hornbuckle BK (2003) Radiometric sensitivity to soil moisture relative to vegetation canopy anisotropy, canopy temperature, and canopy water content at 1.4 GHz. PhD thesis on Electrical Engineering and Atmospheric, Oceanic & Space Sciences, University of Michgan, 2003 18. Adamchuk VI, Hummel JW, Morgan MT et al (2004) On-the-go soil sensors for precision agriculture. Comput Electron Agric 44(1):71–91 19. Wan SS (2001) Fundamentals of soil physics. National Institute for Compilation and Translation, pp 136-145
123
20. Ali MH (2011) Practices of irrigation & on-farm water management, vol 2. Springer Science and Business Media, Berlin 21. Diedrichs AL, Tabacchi G, Gru¨nwaldt et al (2014) Low-power wireless sensor network for frost monitoring in agriculture research. In: IEEE biennial congress of argentina (ARGENCON), Argentina, pp 525-530 22. Le TD, Tan DH (2015) Design and deploy a wireless sensor network for precision agriculture. In: Conference on information and computer science (NICS), 2015 2nd national foundation for science and technology development, Vietnam, pp 294–299 23. Xu J, Zhang J, Zheng X, et al (2015) Wireless sensors in farmland environmental monitoring. In: International conference on cyberenabled distributed computing and knowledge discovery (CyberC), China, pp 372–379 24. Khedo KK, Hosseny MR, Toonah MZ (2014) PotatoSense: a wireless sensor network system for precision agriculture. In: ISTAfrica conference proceedings, Mauritius, pp 1–11 25. Das I, Naveen C, Yadav SS et al (2009) WSN monitoring of weather and crop parameters for possible disease risk evaluation for grape farms–sula vineyards, a case study. In: Proceedings of conference on geospatial technologies and applications, Geomatrix-2009, IIT Bombay, Mumbai, pp 27–29 26. Shinghal K, Noor A, Srivastava N et al (2010) Wireless sensor networks in agriculture: for potato farming. Int J Eng Sci Technol 2(8):3955–3963 27. Haule J, Michael K (2014) Deployment of wireless sensor networks (WSN) in automated irrigation management and scheduling systems: a review. In: Pan African conference on science, computing and telecommunications (PACT), Tanzania, pp 86–91 28. Gutie´rrez J, Villa-Medina JF, Nieto-Garibay A et al (2014) Automated irrigation system using a wireless sensor network and GPRS module. IEEE Trans Instrum Meas 63(1):166–176 29. Ramya V, Palaniappan B, George B (2012) Embedded system for automatic irrigation of cardamom field using Xbee-PRO technology. Int J Comput Appl 53(14):36–43 30. Khriji S, Houssaini DE, Jmal MW et al (2004) Precision irrigation based on wireless sensor network. IET Sci Meas Technol 8(3):98–106 31. Chikankar PB, Mehetre D, Das S (2015) An automatic irrigation system using ZigBee in wireless sensor network. In: International conference on pervasive computing (ICPC), India, pp 1–5 32. Katyara S, Shah MA, Zardari S, Chowdhry BS, Kumar W (2017) WSN based smart control and remote field monitoring of Pakistan’s irrigation system using SCADA applications. Wirel Pers Commun 95(2):491–504 33. Nagarajan G, Minu RI (2017) Wireless soil monitoring sensor for sprinkler irrigation automation system. Wireless personal communications. Springer, Berlin, pp 1–17 34. Zeng XH., Long MS, Qing LIU, Wang XA, Luo WL (2016) Intelligent management system for small gardens based on wireless sensor network. In: International conference on P2P, parallel, grid, cloud and internet computing, Springer International Publishing, Spain, pp 49–56 35. Patel JB, Bhatt CB, Patel B et al (2011) Field irrigation management system using wireless sensor network. In: International conference on engineering, Nirma University, India, pp 1–4 36. Nandurkar SR, Thool VR, Thool RC (2014) Design and development of precision agriculture system using wireless sensor network. In: First international conference on automation, control, energy and systems (ACES), India, pp 1–6 37. Bennis I, Fouchal H, Zytoune O et al (2015) Drip irrigation system using wireless sensor networks. In: Conference on computer science and information systems (FedCSIS), Poland, pp 1297–1302 38. Milind PK, Bhaskar PC (2014) Microcontroller based adaptive irrigation system using WSN for variety crops and development
Neural Computing and Applications
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
of insect avoidance system for better yield. Int J Res Eng Technol (IJRET) 03(07):308–312 Mohandas P, Sangaiah AK, Abraham A, Anni JS (2017) An automated irrigation system based on a low-cost microcontroller for tomato production in South India. In: Computational intelligence in wireless sensor networks, Springer International Publishing, pp 49–71 Panchard J, Rao S, Sheshshayee MS et al (2008) Wireless sensor networking for rain-fed farming decision support. In: Proceedings of the second ACM SIGCOMM workshop on networked systems for developing regions, USA, pp 31–36 Jiber Y, Harroud H, Karmouch A (2011) Precision agriculture monitoring framework based on WSN. In: Wireless communications and mobile computing conference (IWCMC), Turkey, pp 2015–2020 Balendonck J, Hemming J, Van BAJ et al (2008) Sensors and wireless sensor networks for irrigation management under deficit conditions (FLOW-AID). Coordinated by Wageningen University and Research Centre in the Netherlands Yoo S, Kim J, Kim T et al (2007) A2S: automated agriculture system based on WSN. In: Proceedings of ISCE 2007, IEEE international symposium on consumer electronics, USA, pp 20–23 Hemraj S (2014) Power estimation and automation of greenhouse using wireless sensor network. In: 5th international conference– the next generation information technology summit (Confluence), India, pp 436–441 Kassim MRM, Mat I, Harun AN (2014) Wireless sensor network in precision agriculture application. In: International conference on computer, information and telecommunication systems (CITS), South Korea, pp 1–5 Mat I, Kassim MRM, Harun AN (2014) Precision irrigation performance measurement using wireless sensor network. In: Sixth international conference on ubiquitous and future networks (ICUFN), China, pp 54–157 Harun AN, Kassim MRM, Mat I et al (2015) Precision irrigation using wireless sensor network. In: International conference on smart sensors and application (ICSSA), Malaysia, pp 71–75 Punetha D, Mehta V (2014) A wireless approach to real time remote monitoring system examining environmental parameters
49.
50.
51.
52. 53.
54. 55. 56.
57.
58.
59.
60.
using feasibility of a GSM module. In: 2014 IEEE international conference on: in advances in engineering and technology research (ICAETR), India, pp 1–6 Anurag D, Roy S, Bandyopadhyay S (2008) AGRO-SENSE: precision agriculture using sensor-based wireless mesh networks. In: Innovations in NGN: future network and services, 2008. K-INGN 2008. First ITU-T Kaleidoscope Academic Conference, Switzerland, pp 383–388 Kone CT, Hafid A, Boushaba M (2015) Performance management of IEEE 802.15.4 wireless sensor network for precision agriculture. IEEE Sens J 15(10):5734–5747 Maheswararajah S, Halgamuge SK, Dassanayake KB et al (2011) Management of orphaned-nodes in wireless sensor networks for smart irrigation systems. IEEE Trans Signal Process 59(10):4909–4922 Soe KT (2008) Increasing lifetime of target tracking wireless sensor networks. Proc World Acad Sci, Eng Technol 32:410–415 Miller MJ, Vaidya NH (2005) AMAC protocol to reduce sensor network energy consumption using a wakeup radio. IEEE Trans Mob Comput 4(3):228–242 Bar-Shalom Y, Li XR, Kirubarajan T (2001) Estimation with applications to tracking and navigation. Wiley, New York Lewis FL (1986) Optimal estimation. Wiley, New York Ristic B, Arulampalam S, Gordon N (2004) Beyond the Kalman filter: particle filters for tracking applications. Artech House, USA Heinzelman WB, Chandrakasan A, Balakrishnan H (2002) An application-specific protocol architecture for wireless microsensor networks. IEEE Trans Wirel Commun 1(4):660–670 Wang A, Chandrakasan A (2002) Energy-efficient DSPs for wireless sensor networks’. IEEE Trans Signal Process Mag 19(4):68–78 Wang W, Srinivasan V, Wang B et al (2008) Coverage for target localization in wireless sensor networks. IEEE Trans Wirel Commun 7(2):667–676 Perkins CE, Bhagwat P (1994) Highly dynamic destination sequence-vector routing (DSDV) for mobile computers. ACM SIGCOMM Comput Commun Rev 24(4):234–244
123