Abstract
Landslide is a critical natural geological hazard that causes severe damage to property, infrastructure, and humans. In general, some location-specific factors trigger a landslide. Wireless sensor network (WSN) is an enabling technology to monitor most of the parameters associated with these factors. A challenge in landslide monitoring through WSN is that each sensed data item might be critical whereas the underlying wireless communication is often unreliable. In case of landslides, the terrains have irregular shapes, providing harsh conditions for wireless communication thereby more data loss may be expected. This study focuses on the effect of lossy communication in WSN on the efficiency and accuracy of landslide monitoring systems. To this end, collaborative local data analysis is used to enable each node to decide locally whether its sensed data corresponds to a potential event-of-interest. Through extensive simulations, the performance of various landslide prediction and detection models has been evaluated. By and large, the study lets a significant insight into landslide monitoring before implementing a parameterized application for real-world deployment.
1. Introduction
Wireless sensor networks (WSNs) have a vast area of applications ranging from health care to military surveillance systems, monitoring natural disasters: landslide, flood, forest-fire, geographical surveys, object tracking, targeting, and so forth. This study considers its application in landslide monitoring. Different techniques have been used for monitoring the landslides. The use of WSN is a novel approach for monitoring landslides as it can be deployed in the field to collect real-time data.
Landslides cause severe damage to the lives and infrastructure, blocking the natural routes of water, and affecting the local citizens’ economic and social condition. Like other developing countries, in Pakistan, natural disasters cause significant damage. In Bahrain (2014) and in Bajaur (November 2014) areas of Pakistan, land sliding catastrophes have occurred. In January 2010, a rugged landslide occurred in the Hunza valley blocking the Hunza River, causing 19 fatalities and destroying 26 homes. Consequently, the Hunza River was blocked for about five months. Another landslide was triggered on January 22, 2010, and also caused a flood displacing 6000 people. As per statistics of the Information Bulletin 2010, “Pakistan: landslides,” by International Federation of Red Cross and Red Crescents Societies, about the house damage caused by the landslide of 2010 in Attabad and other villages of Hunza Valley were Attabad village: 103, Sarat village: 33, and Salman Abad: 55. A substantial portion of local citizens’ income is spent on rehabilitation, which has a negative impact on GDP growth. The survey in [1] depicts that 41.8% of locals spent 50% of their income in repairing and rehabilitation, 36.8% spent 80% of their incomes, 11% of houses completely were damaged, 28.6% of houses were partially damaged. An early warning system is indispensable which should give enough time for excavation to minimize this damage.
A landslide is a rare event, a short-lived, critical, and sudden occurring geological phenomenon. Moreover, it is a region rather than a site-specific hazard, depending on the regional environment and geological and topographical conditions. It is a complex phenomenon, which is caused by several factors including slope geometry, soil quality, soil moisture, rainfall, aspect, the vegetation index, construction in that region, load on the surface area, and distance from the roads and rivers. The influence of these factors may vary across regions.
We consider landslide monitoring of an area through wireless sensor nodes. These nodes coordinate in time and space upon the detection of an event to establish a consensus. A challenge in such a system is posed by the fact that each data item sensed by a sensor node may be critical in decision-making. On the other hand, WSN offer communication channels which may lose data. Moreover, the terrains which are monitored for potential landslides are irregular in shape and offer harsh conditions, thus increasing the likelihood of data loss. The nodes might not detect all the events that occurred introducing false negatives. Similarly, insignificant events may be reported causing false alarms.
In this study, we use collaborative local data analysis to enable each node to decide locally whether its sensed data corresponds to a potential event-of-interest. Essentially, each node shares its sensed data with its geographical neighbors through low-power wireless communication. We evaluate the performance of various landslide prediction and detection models through comprehensive simulations.
1.1. Literature Review
In this section, the technologies and techniques that are used for monitoring landslides are discussed that can be categorized into three classes:(i)Statistical methods and rating scheme(ii)Remote sensing(iii)Ground-based systems
1.1.1. Statistical Method and Rating Scheme
In [2], numerical weights are assigned to each of the triggering parameters of the landslide. All the ratings are sum up to get an aggregated value used to determine a landslide’s likelihood. The higher the value obtained, the higher would be the probability of landslide. This method can develop “Landslide Hazard Mappings” to identify the potential landslide hazards in the area. It is simple, but it suffers from subjectivity, as experts’ evaluation is based on analytical studies, and personal biases may interfere while assigning ratings to each parameter.
Rare event logistic regression, a statistical method [3], was also used to develop landslide hazard zonation maps. This method compares the area’s statistical data under consideration with the areas that already experienced landslides based on predictions.
1.1.2. Remote Sensing
In remote sensing, the sensors acquire data while deployed at remote locations. The method does not consider intrinsic parameters, which have a high degree of influence on landslides’ occurrence. Besides, these systems are not much helpful in prediction.
(1) SAR, LISA, and LIDAR. Synthetic Aperture Radar (SAR) uses electromagnetic radiation for generating images of the hill slopes. The images must be clear and of high quality to get an accurate and reliable estimation of the landslide detection.
Laser Interferometer Space Antenna (LISA) is a ground-based InSAR system. LISA gets images of the area from the ground, minimizing the weathering effect in getting quality imagery.
Light Detection and Ranging (LIDAR) is a terrestrial, portable scanning device. The slope area is scanned to obtain data that is then corelated to determine any deformation [4]. This technology is suitable in environments where the deformation is determined only by the morphological data; for example, pit mining can be monitored using LIDAR applications.
(2) Global Positioning System (GPS). GPS is also used for locating the areas of active landslides. It is feasible for landslide detection.
(3) Photogrammetric Technique. The photogrammetric technique uses 3D specialized cameras for monitoring ground movement in the mining areas. Cameras are installed at a height to observe the terrain. In imagery-based monitoring systems, the images obtained may be blurred, affecting the landslide detection process. Deep learning-based algorithms help improve the quality of the images [5].
1.1.3. Ground-Based Techniques
In ground-based techniques, the sensing devices directly contact the terrain, acquiring real-time data.(1)Using antennas: antennas are placed at different locations on a terrain. The tilt in the position of the antenna suggests the displacement of the surface. This method is supported by using the “Analytical Hierarchy Process” and GIS. It is useful only to detect mud sliding when it happens, ignoring most of the triggering factors [6].(2)OTDR: Optical Time Domain Reflectometry is used for monitoring debris flow. It uses fiber optic sensors (reflectometer) that detect vibrations generated due to ground movements. Practically, the reflectometer bends when the ground experiences displacement. OTDR is a useful monitoring technique, but it is not cost-effective.(3)WSN-based methods used for land sliding: since its inception, WSN-based monitoring systems for landslides receive significant attention. In the subsequent section, each of the systems that incorporate WSN is analyzed and discussed.
Sheth et al., in their project SENSLIDE [7], implemented a prototype-based prediction model for land sliding. Their study’s primary focus is to have a scalable, fault-tolerant, energy-efficient, and cost-effective system. Data aggregation and duty cycling are used to minimize the amount of data transmission and energy. The election process implements aggregator alteration to have uniform energy diminution. The selection is based on the level of energy a node has. Fault tolerance and cost-effectiveness are achieved by using many inexpensive sensors and considered stress-strain properties of rocks for prediction.
The study of Andreas Terzis et al. [8] focuses on locating the place that observes slip movement and then making an appropriate prediction about land sliding. They tested their model on a laboratory testbed, considering most of the triggering factors for land sliding. It integrates multiple sensors upon rod-shaped columns that are deployed in a grid fashion. This system identifies the slip surface by determining the moved and unmoved columns. This study uses an approach to minimize power consumption, by turning off all other sensors, the only strain gauge is a turn-on. When it observes some movement, only then, it alerts other sensors. However, in this way, the system may miss important data.
In [9], Rehana Raj et al. proposed a clustering and subclustering approach in which node leaders (NL) and cluster heads (CH) are used. The data transmission and energy consumption are reduced using this approach. Whenever a CH fails, the NL will take the failed CH’s responsibilities, improving the fault tolerance. Priority-based sensors placement strategy is used in this work to deploy sensors in different zones of the terrain. Data is periodically sampled and sent even though no new-event occurs, consuming a considerable amount of energy. The clustering is based on the Graph-Partitioning time complex algorithms that require high computations and energy.
The study of Tejaswi [10] proposed clustering and routing protocols for WSN, which embed dynamic-clustering based on energy level and distance from the base station. Specialized nodes having high energy levels would be the best option for minimizing the reelection computation and energy diminution.
In a joint project, Azzam et al. [11] developed a prototype-based model of a monitoring system in a joint project, Sensor-Based Landslide Early Warning System (SLEWS). Solar-power gateways are used to pass data to the base station. He used sensor fusion, a collective decision that reflects the data from all the sensors. In this way, the data is validated, achieving a high degree of accuracy; however, this work focused on detection. In [12], the ultrasonic signals are used to detect the movement of the land. It assigns duty cycles to the sensor nodes. This method achieves high accuracy by attaining high-level synchronization among the nodes; however, sync pulses transmission may fail.
Rosi et al. in [13] described practical experiences highlighting the problems encountered during the deployment phase. They pinpointed the performance differences in simulation, laboratory testbeds, and real environments. In this system, the sensors are placed based on their characteristics, and the bridging nodes are placed in a relatively stable position increasing the network’s life. In-node compression is used to distinguish between the noise and actual data, reducing data transmission and energy consumption. The data of one node is cross-correlated with the data from another node. If both data imply the likelihood of an event, then the data is forwarded for decision-making.
Chang et al. [14] show the effect of distance, weather, and temporal conditions on signal and transmission strength. The environmental-conditions like humidity, temperature, rainy day, and sunny day also affect packet loss.
Ali et al. [15] provide a significant technical and mathematical background for designing early warning. The sensors used in the testbed were moisture, pressure, and flexible bend sensors.
Smarsly et al. [16] proposed the use of software agents for the assessment of data. The node processes data locally and sent alerts to the BS, reducing data transmission. It does not exploit the clustering and aggregation. It provides an online remote monitoring facility. The soil composition affects the landslide process, which needs to be considered in experimentation; however, this study uses sand in their testbed experiment.
The work of Teja et al. [17] followed a stepwise procedure; in the first step, the change in pressure is monitored, if it varies, then, in the second step, geophone is alerted and sensed any vibration generated in the soil, and, in the third step, the tiltmeter calculates any movement occurring. The transmission is done using the IEEE Zigbee standard specification.
In the study [18], “Portrait-Based Disaster Alert System PDAS” is integrated with WSN. It can transmit multimedia messages and GPS coordinates to mobile users, which requires a large amount of power; however, landslide monitoring systems are less likely provided with enough power source.
Shukla et al. [19] developed a prototype model for generating 2-level alerts; the first level alert is based on a single parameter, and the next level is based on the values of multiple parameters when more than one sensor reaches the threshold value. This study considered most of the triggering parameters that influence slope stability.
Ramesh et al. published successive papers [20–23] regarding landslide prediction on the real-world deployed project. She described the network requirements [24] and selection and placement of sensors, along with protocols specifications. She also shared her practical experiences of the field deployment [22–24]. The data is sent through UDP, including lost packets’ recovery and secure transmission. The same author in her work [20] uses a consensus approach; the readings from the same type of sensors can be combined and corelated to remove redundancy from the data.
Zhen et al.’s [25] work gives the prospects of integrating the sensing devices with the Low Earth Orbit (LEO) satellites using the 6G communication technology. This would make the data collection more comfortable and would also minimize the complexity in routing the packets to the base stations.
2. Experimental Setup
2.1. System Model
We consider a decentralized monitoring system consisting of sensor column nodes installed in the area to be monitored. The concept of a physical sensor column [23] is assumed, upon which different sensors are mounted. Due to its decentralized nature, each node acts as a base station (BS) for its neighbor nodes. Each node periodically senses the data in its proximity, records that data in its local buffer, and shares it with neighbors through a low-power wireless communication medium. The broadcast data is received and locally stored by the neighbors; each applies aggregation to get a spatial summary of the data. The aggregation is essentially a landslide prediction model to predict the landslide. A similar setup can also be used for detecting the landslide. This system is simulated through three different topologies: grid, random, and triangular, as depicted in Figure 1.
2.2. System Parameters
The landslide is a complex phenomenon and is affected by various factors. In this study, rainfall is considered as a driving factor. To this end, primary parameters like pore water pressure, stress, displacement, and the vibrations and derived parameters like antecedent rainfall and rainfall intensity have been considered.
2.3. Performance Metrics
The landslide monitoring system ought to be efficient and accurate. Thus, the key performance metrics chosen for this system are efficiency and accuracy. The number of false alarms determines the system’s efficiency. A false positive (FP) is generated when the system triggers and no event-of-interest has occurred. When an event-of-interest is missed, the system is said to have generated a false negative (FN).
2.4. The Aggregation Models
The rain gauge mounted on a sensor column records the hourly event rainfall. In the following, a brief description of various aggregation models is given.
The intensity duration model is based on the rainfall intensity and duration mapped onto an intensity duration model [26]. This model dynamically determines the threshold for the event rainfall intensity and determines the likelihood of a landslide. Mathematically, the model can be expressed aswhere “I” is event intensity in mm/hr and D is the event rainfall duration in hours. The original intensity duration model derived in [26] has 14.82 and −0.39 as the and β values, respectively. However, several studies then derived region-specific values depending on geological characteristics that the field-specific expert can best define. This system can be configured depending upon the region under consideration. For now, in simulation, the alpha (α) as 91.46 and beta (β) as −0.82 are assumed [27].
The antecedent rainfall, if any, considerably affects the current event rainfall threshold. Antecedent rainfall lowers down the event rainfall threshold required for triggering landslides. Different studies consider different antecedent days of rainfall, For example, 10, 5, or 2 days. Following [28], we assume the five-day antecedent rainfall. The antecedent rainfall and event rainfall are then mapped into a function, which determines the event rainfall threshold:
The function requires the parameters sigma (σ) and gamma (γ) to be configured. We assume their values as 80.7 and 0.1981, respectively [28].
The rainwater enters the rocks’ cracks and pores, exerting pressure on these rocks and ultimately causing the rocks to break, resulting in the displacement of rocks and soil. The displacement transducers are used to measure these displacements. The geophone is used to support the displacement transducers and measure the vibrations in soil caused by the displacement of soil layers.
The rainwater in the soil pores also generates pressure called the pore water pressure (PWP). The pore pressure is measured through a piezometer. The pore pressure also has a considerable share in the occurrence of land sliding. Pore pressure values greater than 7 and up to 9.4 kPa have greater chances of sliding.
2.5. Validation
For validating our approach, we compare our decentralized approach’s performance using unreliable wireless communication with an ideal counterpart, characterized by reliable communication. There is no packet loss in an ideal situation; while considering wireless communication, the system may experience packet loss.
3. Results and Discussion
In geospatial networks, the nodes gather location-sensitive data from within particular proximity. Monitoring large areas of land requires many wireless nodes, thus increasing the network size. On the other hand, improving the quality of data collected, relatively dense deployment of the network nodes may be desirable. Therefore, the system is evaluated with respect to both parameters, namely, network size and network density.
3.1. Packet Loss
For this specific application, a landslide monitoring system, the relevant events are rare and critical. So, in principle, the system cannot afford to lose any of them.
3.1.1. Effect of Network Size on Packet Loss
The parameters like soil moisture, pore pressure, displacement, and vibration are location-specific; that is, they vary in almost 100 m distance [7]. To record relevant data, the sensors monitoring these parameters need to be increased to get location-specific data. We increment network size across various experiments to simulate its effect on packet loss. More specifically, we simulated ten different network sizes using wireless network topologies.
It is evident from the graph in Figure 2 that the packet loss does not increase significantly with the increase in network size. The packet loss in wired topology is zero because wired communication is considered reliable compared to wireless communication. In this study, the wired topology acts as a benchmark for the wireless communication topologies. The results show that the lowest packet loss has been observed in the random topology, indicating its suitability for landslide monitoring. Moreover, this topology may be more convenient to deploy considering the irregular structure of the terrain.
3.1.2. Effect of Density on Packet Loss
The land to be monitored is usually a vast area where each region may have its characteristics. The nodes’ density may be a crucial parameter to collect more analytical data for all these regions. A dense network gives highly grain data, which is desirable. However, high density may degrade the quality of communication among the nodes due to high contentions and high packet loss. The system is examined with respect to density keeping the topology’s shape the same as depicted in Figure 3. The effect of density on packet loss is evaluated to analyze the overall performance of the system. Random topology, as compared to the other two wireless topologies, experiences low packet loss. The network sustains an increase in density to a certain extent. However, a further increase in density raises the packet loss. Moreover, random topology endures more extended than the other two, as shown in Figure 4.
3.2. Yield in Consensus Development
Yield is the degree of participation by the neighbor nodes in data aggregation. In this section, the effect of network size and density on the yield is examined.(1)Effect of network size on yield: ideally, the yield may increase with increasing the network size; however, this is not always the case; specifically, in wireless communication. A decentralized or distributed approach is one of the means of keeping the yield optimum in wireless communication. Results depict that due to the application’s decentralized nature, the network size does not affect the yield significantly. As shown in Figure 5, wired topology being reliable observes the highest yield.(2)Network density versus yield: in dense networks, the yield is expected to be degraded due to interference and high contentions. Similarly, the yield also starts decreasing with decreasing density (sparse networks); that is, from scenario eight and onward, the yield starts decreasing as shown in Figure 6, potentially due to an increase in internode distance. This indicates that optimum placement of sensor nodes is needed to get a higher level of yield.
3.3. Performance of Monitoring Models
For prediction, the intensity versus duration, antecedent rainfall, and pore pressure based models are considered in this study. The accuracy (ratio of true positives plus true negatives to the total number of events), false positives, and false negatives of these models are evaluated.(1)Intensity duration (ID) model: This model’s accuracy is simulated and observed with respect to varying the size and density of the network. As rain falls on a region in an approximately uniform way, each node has almost the same rainfall value. Therefore, its value is not critical in terms of message loss. Hence, the aggregation may not be affected significantly. Due to distributed nature of this system, the increase in network size does not significantly affect packet loss and eventually the accuracy, as shown in Figure 7. However, varying network density significantly impacts this model’s accuracy, as depicted in Figure 8. False positives and false negatives are also examined with varying network size and density, as shown in Figures 9 to 11.(2)Antecedent rainfall model: This model requires twofold data, the event rainfall, and the 5-day previously recorded rainfall data. This model’s performance is also examined in terms of the false alarms and accuracy with varying network size and density. Figures 12 to 16 show the accuracy, FPs, and FNs with respect to size and FPs with respect to density.(3)PWP-based prediction model: The region-specific parameter PWP has distinct values in different regions. Its values in the slope’s crown region are much lower than in the middle and toe regions due to the frequent downward flow of water. Hence, this parameter is critical and should be considered. It has a significant effect on the aggregation process and eventually affects the prediction. We evaluate with various sizes and densities of the network. The PWP-based prediction is far behind the ID prediction model due to its region-specific parametric values in terms of accuracy and false alarms, as shown in Figures 17 to 21.(4)Detection model: The landslide detection model is far more critical and essential, as it provides less time for evacuation. The parameters considered in this study for detection are region-specific, making these values more significant, affecting the consensus and detection process.
The accuracy of this model with respect to size and density is shown in Figures 22 and 23. The performance of this model in terms of FPs and FNs is lower than the prediction models.
However, this is due to the nature of the parameters involved in the model. The parameters like displacement and stress are area-specific. Usually, the crown region of the slope experiences more stress due to gravity and has high values of displacement. One point is experiencing more stress, while the other has zero stress values. The same is the case of displacement. In such situations, the consensus process is greatly affected by communication failures and detection processes, resulting in false alarms and ultimately questioning the system’s credibility employing wireless communication protocols. Increasing the network size does not significantly affect the number of FPs and FNs, as depicted in Figures 24 and 25. However, in dense networks, the system generated a significant number of false alarms, as shown in Figures 26 and 27.
3.4. Comparison of Monitoring Models
For a landslide monitoring system, a trusted model may be employed to predict and detect the landslide well enough to provide ample time for evacuation. For this purpose, these models are compared. The prediction is made using the well-known intensity duration model, the antecedent rainfall model, and the PWP-based model.
Due to the generalization of the rainfall values, ID and antecedent rainfall models are expected to be performing well by setting the appropriate configurations.
The prediction model using the PWP as a driving factor makes it vulnerable to false alarms (FPs and FNs). Similarly, the detection model using the stress, displacement, and vibration in soil layers is region-specific to a great extent, making the detection process more vulnerable to false alarms.
To bolster this statement with scientific support and develop a clear design idea, these models’ efficacy is examined in this study, with both varying network size and densities. Figure 28 depicts the performance comparison of the different monitoring models in terms of accuracy.
The percentage of FPs and FNs generated by these models with varying network size and density is compared in Figures 29 to 32. These graphs represent that the model using more generalized parameters is experiencing a less number of false alarms (FPs and FNs). The ID model shows better results as compared to other monitoring models, in terms of FPs and FNs, as shown in Figures 30 and 31. The reason is that rainfall equally affects the different terrain regions, and its packet loss in data dissemination has no significant effect.
4. Conclusion and Future Work
Landslide is a rare and sudden event. At the same time, it is a critical hazard which causes damage to both infrastructure and human life. Slopes in the vicinity of infrastructure and humans need to be monitored by a vigilant and accurate system to avoid such damage.
As landslide is affected by several factors from both internal and external environments, the monitoring and prediction systems need to cover all those possible factors that can influence slope stability and trigger a landslide. These parameters need to be monitored and calculated in real time; changes in these parameters’ values dictate the likelihood of the landslide occurrence. The prediction should be early enough to excavate timely from the area of the exhibit.
We considered a wireless sensor network to monitor these parameters. The performance of various landslide prediction and detection models which were dependent on these parameters was evaluated in a lossy communication environment of the WSN.
We have seen that the network size has no significant effect on packet loss, indicating the scalability of our approach. However, the network sustains performance for a specific range of densities. Outside that range in density, the packet loss rises. This means that deciding on the network density requires special attention. The yield, which influences the quality of decision about an event, sensed by a node, has shown a similar tendency for packet loss as was exhibited by the network size and density.
The intensity versus duration model showed no significant change in accuracy with lossy communication. This can be attributed to the observation that rainfall is a regional parameter. The antecedent rainfall model is not influenced by the network size; however, it shows a decrease in accuracy with a decrease in network density. This behavior can be attributed to the increase in communication loss due to network sparsity. With regard to the effect of network size and density on accuracy, the pore water pressure model showed similar behavior to the antecedent rainfall model.
The accuracy of the detection model is influenced by network size and density. Like the prediction models, the accuracy is more sensitive to the network density than the network size. Moreover, the overall accuracy level of the detection model is lower than the prediction models. This can be attributed to the fact that prediction models were mainly using the rainfall parameter which is usually uniform across the region. On the other hand, the detection model is using parameters such as displacement and stress which are site- rather than node-specific, thereby, making the model more sensitive to communication failures in the network.
We are led to the conclusion that the models which use region-wide parameters are producing a smaller number of false alarms (FPs and FNs). However, further study is needed to come up with an intricate procedure that exploits both the region- and site-specific parameters to improve the accuracy of the landslide monitoring system using wireless sensor networks.
4.1. Future Work
We plan to implement the proposed system on a testbed of 4 nodes using the LoRa wireless communication, which has advantages over Zigbee in certain aspects, including transmission range and low-power consumption. We also consider integrating the wireless sensors with web-based services and machine learning to adapt to region-specific configurations during landslide events monitoring.
Data Availability
The data used to support the findings of this study will be made available on Github.
Conflicts of Interest
The authors declare that there are no conflicts of interest regarding the publication of this paper.
Acknowledgments
The authors would like to thank the Deanship of Scientific Research at Prince Sattam Bin Abdulaziz University Al-Kharj Saudi Arabia for supporting the research. The authors would also like to thank the University of Engineering and Technology, Peshawar, Pakistan, for providing the equipment for experiments in this research.