A dynamic distributed boundary node detection algorithm for management zone delineation in Precision Agriculture

Abstract

Dividing a larger area into smaller subregions is a well addressed problem in Precision Agriculture (PA) where the existing application specific solutions (laboratory based) require human intervention and result in static region demarcation schemes. However, the boundary of a subregion is subject to change with various soil and environmental parameters. On the other hand, Wireless Sensor Networks (WSNs), a potential candidate to produce dynamic subregions due to its real-time decision making capability, are utilized as merely a data collection unit for PA.

In an attempt to introduce the in-network decision making feature of WSN in PA, design of a novel three layered WSN-CPS architecture is presented in this work and at layer-I the distributed mechanism for region demarcation is proposed. The proposed scheme identifies nodes based on data values and position information, that serve as the boundary of a subregion and data transmitters to the base station for final decision making. Existing methods for boundary node detection identify network boundary nodes (not designed to demarcate the interior boundary), coverage hole boundary nodes (identifies boundary nodes of a smaller portion in network), and event boundary nodes (not able to identify outer boundary). The proposed work identifies network boundary nodes, coverage hole boundary nodes, and the subregion boundary nodes accurately. It takes various critical situations into account while labelling of nodes and shows its reliability in node failure conditions. Impact of varying transmission range and number of nodes is analyzed on proposed mechanism via simulation. In a comparative study with recent network boundary node detection scheme, decrease of 32% and 30% is seen at 200 and 300 nodes respectively in terms of energy consumption.

Introduction

Cyber Physical Systems (CPSs) are being introduced as a solution to future smart infrastructures. Main objective of CPS is to integrate the computation and communication systems with the physical world being sensed and controlled by sensors and actuators respectively to facilitate high degree of automation and real-time decision making in large scale systems (Zeadally and Jabeur, 2016). Wireless Sensor Network (WSN) is a basic element in realization of CPS and plays a vital role that is not limited merely to data forwarding. Wide spread application domain of WSN-CPS, where it facilitates the self operability and system autonomy towards the accomplishment of collective goals, includes Precision Agriculture (PA) and Structural Health Monitoring (SHM).

PA is all about taking within field variability into account and incorporating that variability into management decisions (Haghverdi et al., 2015). Within field variability depicts the heterogeneity of parameters like soil moisture, temperature, Available Water Content (AWC), mineral content that can be the result of spatial and temporal variations of climate, topography, and biological activities (Córdoba et al., 2013). PA specific WSNs are deployed for Variable Rate Fertilization (VRF), Variable Rate Irrigation (VRI), cattle monitoring, etc. WSN-CPS architecture to monitor the field condition is proposed in (Chen et al., 2015). Four layered design is capable of acquiring data from physical sensor layer and processing it with the help of six web services at higher layers. Another three layered architecture is proposed in (Nie et al., 2014) consisting of physical, network, and decision layer. Sensors (physical layer) transmit data to switches, routers (network layer) that forwards it to central server for data processing. Authors in (Stark et al., 2013) have designed an algorithm for efficient data acquisition by robots. These PA specific WSN-CPS architectures are facing two main challenges: first is the full fledged architecture design of CPS and second is the exploitation of in-network decision making paradigm of WSN. In our work, we propose three layered WSN-CPS architecture named as AgriCPS (architecture is discussed later in this section) for VRI and design its layer-I leveraging in-network computing leaving higher layer design for future work.

VRI is irrigating different regions according the requirement of those regions. In the state-of-art VRI systems (O'Shaughnessy et al., 2015), (Nikolidakis et al., 2015), (Haghverdi et al., 2015), and (Zhang et al., 2013) different subregions of a large area, named as Management Zones (MZs) are predetermined and remain fixed throughout. However, the concern is, the predetermined regions' physical parameters can vary with respect to time and space. The most prominent methods for MZ delineation namely, Yield maps, satellite images, and on-the-go sensors (sensors mounted on vehicle) do not account for this aspect. Recent studies from agronomy ((Sudha et al., 2011), (O'Shaughnessy et al., 2015), (Vellidis et al., 2008), and (Haghverdi et al., 2015)) report the need for identifying MZs dynamically. This necessitates estimation of MZs dynamically based on physical parameters.

This paper aims at developing a distributed and energy efficient method to tackle region demarcation problem that accomplish layer-I goals of AgriCPS.

Region demarcation is a well-researched problem in WSN and the existing mechanisms can be classified on the basis of approaches used and applications (Classification Tree is shown in Fig. 1). Based on various applications, region demarcation can be broadly classified into physical and virtual boundary identifications. Physical demarcation of a region is the partitioning of an area on maps or through physical boundary. It is further divided into MZ delineation and Sub Structure (SS) formation. For VRI and VRF applications, farmland is divided into MZs in form of maps. On the basis of approaches used MZ delineation is classified into on-the-go sensors (Viani et al., 2017), (Gili et al., 2017), yield maps, and satellite images that produce static MZs. For better monitoring and maintenance purpose a structure in SHM is subdivided into smaller areas called SSs with the help of wired infrastructure (Bhuiyan et al., 2016), (Hackmann et al., 2013).

Other counterpart of physical boundary identification is the virtual boundary identification that groups sensors according to application. It can be further classified into entire and partial region demarcation schemes. Problems such as routing, load balancing, scalability, etc. require entire region to be demarcated into smaller subregions (entire region demarcation) whereas target tracking, continuous object boundary detection, detecting energy holes require a part of the region to be located (partial region demarcation). Applications pertaining to partial region demarcation is grouped into network boundary identification, coverage hole boundary identification, and event region boundary identification. Network boundary identification is locating perimeter of a region by identifying boundary nodes laying on it whereas coverage hole boundary identification is marking a smaller region that is left unmonitored. Boundary of network and coverage hole can be demarcated by utilizing position information (geometrical) (Zhang et al., 2009), (Wu et al., 2007), (Fang et al., 2006), (Qiu and Shen, 2014), (Wang et al., 2006a), connectivity information (topological) (Wang et al., 2006b), (Khan et al., 2011), (Hsieh and Sheu, 2009), (Dabba and Beghdad, 2014), (Saukh et al., 2010), (Beghdad and Lamraoui, 2016), (Bejerano, 2011), (De Silva and Ghrist, 2007, (Funke, 2005), (Li and Zhang, 2015) and distribution assumptions (statistical) (Fekete et al., 2005), (Bi et al., 2006), (Chen et al., 2012),(Fekete et al., 2004),(Tahbaz-Salehi and Jadbabaie, 2010) of nodes in the network. Event region demarcation locates the position of an event occurring in network. The approaches followed in this subclass are event boundary nodes detection (Ding and Cheng, 2009), (Liao et al., 2004), (Kundu et al., 2018a), (Imran and Ko, 2017), (Duh et al., 2013), (Ren et al., 2008), (Dogandzic and Zhang, 2006), (Chu and Ssu, 2014), (Shukla et al., 2017), (Ding et al., 2005) and various machine learning algorithms including Support Vector Machine (SVM) (Sakaki et al., 2010), (Ould-Ahmed-Vall et al., 2011), classification (Li et al., 2009), (Zhu et al., 2007), (Xue et al., 2006), and Bayes’ theorem (Tang et al., 2012), (Liu et al., 2016). On the basis of approaches used entire region demarcation is classified into cluster formation (Le et al., 2008), (Yi et al., 2007), (Gong et al., 2008), (Youssef et al., 2006), (Demirbas et al., 2004), (Lin and Gerla, 1997), (Manjeshwar and Agrawal, 2002), (Xu et al., 2014a), (Xu et al., 2014b), (Luo et al., 2005), (Jung et al., 2007), (Lindsey et al., 2002) and Voronoi diagrams (Sharifzadeh and Shahabi, 2004), (Bash and Desnoyers, 2007), (He et al., 2018), (Zhou et al., 2009) that partition entire region into smaller subregions based on distance among sensor nodes.

Problem of MZ delineation requires entire farmland to be divided into smaller subregions where implemented techniques must take into account dynamicity of subregions based on sensed parameters. In classical WSN, existing clustering (Le et al., 2008), (Lindsey et al., 2002), etc. and Voronoi partitioning mechanisms (Sharifzadeh and Shahabi, 2004), ( Zhou et al., 2009), etc. segment entire network into smaller groups of sensors on the basis of distance among sensor nodes. As a result, identified subregions remain unaffected by spatio temporal variations in sensed values and can be referred as static mechanisms with respect to sensed values. Since VRI requires dynamic MZ delineation method, these mechanisms fulfill partial requirement of problem at hand. Existing mechanisms of partial region demarcation (Zhang et al., 2009), (Liu et al., 2016), etc. are designed to identify the boundary of a smaller part in the network. Thus, these approaches are not suitable for MZ delineation. Apart from meeting application specific requirements energy efficiency and accuracy are the two important design considerations of region demarcation algorithms in WSN. For entire region demarcation, centralized Voronoi partitioning methods are energy consuming and distributed approach-es (Sharifzadeh and Shahabi, 2004), (Bash and Desnoyers, 2007), and (Zhou et al., 2009) are unable to construct exact partitions having insufficient information (He et al., 2018). Another method in this area is clustering that is a two phase process namely, Cluster Head (CH) election and cluster formation. CH election can be centralized (decision taken by Base Station (BS)) and distributed (sensors decide locally). Centralized approaches achieve global optimal solution but consume more energy whereas distributed approaches provide local optimal solution and are energy efficient (Xu et al., 2017). The chosen CH advertises itself and neighboring nodes join its cluster depending on various parameters like communication cost, hop count, physical distance, and size of clusters. Distributed approaches are energy efficient in comparison to centralized ones but communication oriented CH election and cluster formation phases consume considerable amount of energy (Tyagi and Kumar, 2013), (Liu, 2012), (Afsar and Tayarani-N, 2014). Thus, centralized algorithms are energy consuming and distributed approaches ensure low energy consumption at the cost of accuracy.

Upon analyzing the existing region demarcation schemes and in light of MZ delineation requirements for VRI the following research questions arise:

• Will the existing partial region demarcation (event/coverage hole/network boundary) approaches be able to demarcate MZs in VRI?

• Instead of applying two different mechanisms, can a single and energy efficient method detect boundary nodes of a subregion and network as well?

• Will the technique result in subregions that reflect spatial and temporal variations of soil parameters?

The research questions lead to the design of a novel application layer mechanism that formulates region demarcation problem as detecting nodes laying on the boundary of a subregion. For example, in Fig. 2 an area A is partitioned into three smaller regions a₁, a₂, a₃ where nodes laying on the boundary (solid circles) of a subregion are Boundary Nodes (BNs) and others are Interior Nodes (INs) for the subregion.

Data dissimilarity and position with respect to its one-hop neighbors are the two criteria for deciding if a node is BN or IN. A node is termed as IN if data dissimilarities among one-hop neighbors are less than a predefined threshold and as BN otherwise. This process faces two main challenges. First, a node laying on network boundary and having similar data to its neighbors, may mark itself as IN. For example, node a₁₁ may have data similar to a₁₂, a₁₃, and a₁₄ (see Fig. 2) and in this situation it will mark itself as IN. However, it is a node laying on the network boundary. The nodes laying on boundary of the network is termed as Network Boundary Node (NBN). Detecting NBNs is important for following reason - BS requires information about soil properties for decision making. In addition, it will need boundary information of a subregion. Why this piece of information is required? Consider a situation where actuator (sprinkler) lies just above boundary of a subregion. Amount of water required, is different for either sides of the boundary. Secondly, percolation property of soil effects water retained by soil. So, it may happen that amount of water absorbed by soil is different for the sides of boundary. To incorporate these aspects into decision, complete picture of boundary is required. NBNs along with Interior BNs (IBNs) transmit data values and coordinates to BS. Depending on these information BS can control the actuator (tilt or put sprinkles in off mode). Second issue with data dissimilarity criteria is the inability to detect coverage holes and micro holes. Micro holes are convex polygons (may not always correspond to a coverage hole) appearing in network modeled as graph where nodes are vertices and connection among them are edges. Micro holes are termed as voids in this work.

Thus, this approach is found to be ineffective in identifying NBNs and voids. Hence, we introduced a position dependent criteria for accurate labelling of nodes. Rationale behind this criteria is explained with the help of Fig. 3.

A node is an IN (with respect to position) if all of its 1-hop neighbors form a virtual enclosure (refer Fig. 3a). If this enclosure is broken somewhere (Fig. 3b), the node can lie either on boundary of a void or network. To decide if a node i is NBN or void BN (VBN), we implement CNF-REP exchange process (discussed in Section 4.1). VBNs are further classified as INs or IBNs depending on data dissimilarity criteria. IBNs are the nodes laying on boundary of a subregion that are not NBNs (for example, a₁₅ (in Fig. 2)). Thus, an in-network mechanism is proposed that works in distributed fashion, incorporating data and position information to classify nodes as IN (a₁₃), IBN (a₁₅), and NBN (a₁₁).

Above mentioned distributed procedure is followed on layer-I of AgriCPS. Different layers of AgriCPS along with their functionality is shown in Fig. 4. The inter-sensor communication based mechanism is initiated when data sensed by a sensor cross a predefined threshold and labels nodes locally based on data and position information. On layer-II communication between sensor and the irrigating machinery Linear that acts as Mobile BS (MBS), takes place. Decision making based on gathered data and generation of control actions (irrigation amount and time) are layer-III tasks that is accomplished by MBS. According to these control actions, subregions are irrigated by sprinklers (actuators).

Boundary node identification mechanism has been widely used for event region detection (Ding and Cheng, 2009), (Ding et al., 2005), etc. , coverage hole problem (Zhang et al., 2009), (Qiu and Shen, 2014) etc., and network boundary detection (Wang et al., 2006b), (Hsieh and Sheu, 2009), (Saukh et al., 2010), (Fekete et al., 2005), etc. The solutions pertaining to event region detection rely mainly on data comparison (Wu et al., 2019) and fusion of neighbors to detect event region boundary. NBNs may have similar data values to its neighbors (a₁₁ in Fig. 2) and hence can be a subject of wrong labelling. EEBD (Wu et al., 2019), a recent event detection scheme identifies 60% NBNs (refer Table 1). Algorithms finding coverage hole (Zhang et al., 2009), (Qiu and Shen, 2014), etc. are designed to locate BNs of an area that is not monitored by any sensor. These algorithms are not able to find NBNs and micro holes. The network boundary algorithms (Wang et al., 2006b), (Hsieh and Sheu, 2009), etc. only find NBNs and other boundary detection algorithms (Zhang et al., 2009), (Wu et al., 2007), (Wang et al., 2006a) are unable to differentiate among VBNs and NBNs. Identifying IBNs along with NBNs and efficient utilization of resources are the applicative scenario's requirements. None of the solutions make a best fit for these requirements. Hence, a lightweight distributed mechanism for region demarcation is proposed with following major contributions:

• A distributed mechanism to label the nodes as INs, IBNs, and NBNs is proposed. NBNs demarcate the network boundary while IBNs are nodes laying on interior boundary of a subregion (MZ). During the labelling process VBNs are also detected as a by product with a clear distinction with NBNs.

• The distributed process takes various critical situations into account, ignored by most of the related work. Incorporating these situations into the proposed solution lead to accurate labelling of nodes.

• Reliability of proposed mechanism in node failure conditions is proved with the help of various cases and at the end, limitation of the algorithm is pointed out after analyzing the performance in various network settings.