A big challenge in terms or landslide risk mitigation is represented by increasing the resiliency of society exposed to the risk. Among the possible strategies with which to reach this goal, there is the implementation of early warning systems. This paper describes a procedure to improve early warning activities in areas affected by high landslide risk, such as those classified as critical infrastructures for their central role in society.
This research is part of the project “LEWIS (Landslides Early Warning Integrated System): An Integrated System for Landslide Monitoring, Early Warning and Risk Mitigation along Lifelines”.
LEWIS is composed of a susceptibility assessment methodology providing information for single points and areal monitoring systems, a data transmission network and a data collecting and processing center (DCPC), where readings from all monitoring systems and mathematical models converge and which sets the basis for warning and intervention activities.
The aim of this paper is to show how logistic issues linked to advanced monitoring techniques, such as big data transfer and storing, can be dealt with compatibly with an early warning system. Therefore, we focus on the interaction between an areal monitoring tool (a ground-based interferometric radar) and the DCPC. By converting complex data into ASCII strings and through appropriate data cropping and average, and by implementing an algorithm for line-of-sight correction, we managed to reduce the data daily output without compromising the capability for performing.
Urbanization, especially in mountain areas, can be considered a major cause for high landslide risk because of the increased exposure of elements at risk. Among the elements at risk, important communication routes, such as highways, can be classified as critical infrastructures (CIs), since their rupture can cause chain effects with catastrophic damages on society (Geertsema et al., 2009; Kadri et al., 2014). On the other hand, modern society is more and more dependent on CIs and their continuous efficiency (Lebaka et al., 2016), and this has increased their value over the years. The result is a higher social vulnerability in the face of loss of continuous operation (Kröger, 2008). The main objective was to improve the social preparedness for the growing landslide risk, according to the suggestions of several authors (Gene Corley et al., 1998; Baldridge and Marshall, 2011; Urlainis et al., 2014, 2015). This led to the development of several approaches and frameworks for increasing the resiliency of society exposed to the risk (Kröger, 2008; Cagno et al., 2011, and references therein). The resiliency policy involves not only prevention activities but also, and more importantly, those activities needed to maintain functionality after disruption (Snyder and Burns, 2009) and to promptly alert people of incoming catastrophes in order to protect them and prepare for a possible damaging of the endangered CI. Among these activities, the implementation of integrated landslides early warning systems (i.e., LEWIS: Landslides Early Warning Integrated System: An Integrated System for Landslide Monitoring, Early Warning and Risk Mitigation along Lifelines; Versace et al., 2012; Costanzo et al., 2016) reveals its increasing importance.
In this context, the methodology described in this paper has been conceived; it has been tested and validated on a portion of an Italian highway that is affected by landslides and that was selected as a case study: it is located in southern Italy, along a section of the A16 highway, an important communication route that connects Naples to Bari, where a ground-based interferometric synthetic aperture radar (GB-InSAR) has been installed at the test site in order to obtain spatial monitoring data.
One of the main drawbacks of advanced instruments such as GB-InSAR is how to handle the large data flow deriving from continuous real-time monitoring. The issue is to reduce the capacity needed for analyzing, transmitting and storing big data without losing important information. The main feature of this paper is indeed the management of monitoring data in order to filter, correct, transfer and access them compatibly with the needs of an early warning system.
The GB-InSAR is composed of a microwave transceiver mounted on a linear rail (Tarchi et al., 1997; Rudolf et al., 1999; Tarchi et al., 1999). The system used is based on a continuous-wave–stepped-frequency radar, which moves along the rail at millimeter steps in order to perform the synthetic aperture: the longer the rail, the higher the cross-range resolution. The microwave transmitter produces, step by step, continuous waves around a central frequency, which influences the cross-range resolution and determines the interferometric sensitivity, i.e., the minimum measurable displacement, usually largely smaller than the corresponding wavelength.
The radar produces complex radar images containing the information relative to both phase and amplitude of the microwave signal backscattered by the target (Bamler and Hartl, 1998; Antonello et al., 2004). The amplitude of a single image provides the radar reflectivity of the scenario at a given time, while the phase of a single image is not usable. The technique that enables to retrieve displacement information is called interferometry and requires the phase from two images. In this way, it is possible to elaborate a displacement map relative to the elapsed time between the two acquisitions.
The main added value of GB-InSAR is its capability to blend the boundary between mapping and monitoring, by computing 2-D displacement maps in near-real time. The use of this tool to monitor structures, landslides, volcanoes and sinkholes is widely documented (Calvari et al., 2016; Di Traglia, 2014; Intrieri et al., 2015; Bardi et al., 2016, 2017; Martino and Mazzanti, 2014; Severin et al., 2014; Tapete et al., 2013), as well as for early warning and forecasting (Intrieri et al., 2012; Carlà et al., 2016a, b; Lombardi et al., 2016).
GB-InSAR systems probably reveal their full potential in emergency conditions. They are transportable and only require from a few tens of minutes to a few hours to be installed (depending on the logistics of the site). Moreover, they can detect “near-real-time” area displacements, without accessing the unstable area, 24 h a day and in all weather conditions (Del Ventisette et al., 2011; Luzi, 2010; Monserrat et al., 2014). On the other hand, some limitations reduce the GB-InSAR technique applicability: first of all, the scenario must present specific characteristics in order to reflect microwave radiations, maintaining high coherence values (Luzi, 2010; Monserrat et al., 2014); only a component of the real displacement vector can be identified (i.e., the component parallel to the sensor's line of sight); and maximum detectable velocities are connected to the time that the system needs to obtain two subsequent acquisitions. Sensors need power supply that, for long-term monitoring, cannot be replaced by batteries, generators or solar panels.
With the specific aim of performing the function of an early warning system, data acquired in situ must be sent automatically to a “control center” where they are integrated into a complete early warning system procedure (Intrieri et al., 2013). In this sense, another main limitation is represented by the necessity to transfer a high quantity of data, whose weight has to be reduced to the minimum, in order to reduce the load on transmission network.
The employed system is a portable device designed and implemented by the Joint Research Center (JRC) of the European Commission and its spin-off company Ellegi-LiSALab (Tarchi et al., 2003; Antonello et al., 2004).
Morphological features, hydrogeological factors and sudden rainfall can cause diverse types of movements or fall of earthy and rock materials. The unpredictability and diversity of these events make structural interventions often inappropriate to reduce the related risk and a real-time monitoring network difficult to implement.
In the last decade, wireless sensor networks (WSNs) have been used extensively in various fields. A significant increase in the use of WSNs – due to their simplicity and the low cost of installation, manufacturing and maintenance – has been recorded in the framework of environmental monitoring applications (Intrieri et al., 2012; Liu et al., 2007; Yoo et al., 2007). Distinct types of sensor nodes of these networks, distributed with high density in the monitored areas, send environmental information to the concentrators nodes, generating a considerable amount and a wide variety of collected data. Due to the significant growth of data volumes to be transferred, the WSNs require flexible ad hoc protocols able to respect constraints related to energy consumption management (Hadadian and Kavian, 2016; Khaday et al., 2015; Parthasarathy et al., 2015). In particular, many protocols have been developed that offer data aggregation patterns to optimize the sensor nodes' battery life (Kim et al., 2015) or sleep–measurement–data transfer cycles to minimize the energy consumption (Fei et al., 2013; Venkateswaran and Kennedy, 2013).
LEWIS (Costanzo et al., 2016) uses heterogeneous sensors, distributed in the risk areas, to monitor the several physical quantities related to landslides. The measured data, through a telecommunications network, flow into the data collecting and processing center (DCPC), where, using suitable mathematical models for the monitored site, the risk is evaluated and eventually the state of alert for mitigation action is released (Fig. 1).
The system, through a modular architecture exploiting a telecommunication network (called LEWARnet) based on an ad hoc communication protocol and an adaptive middleware, has a high flexibility, which allows for the use of different interchangeable technological solutions to monitor the parameters of interest.
LEWIS architecture.
The network has been equipped with both single-point sensors and area sensors. The present paper addresses a sub-network comprising an area sensor, the GB-InSAR.
The different sensors types generate asynchronous traffic, thus imposing the adoption of an ad hoc transmission protocol. This can support an asynchronous transmission mode to the DCPC, and it is equipped with message queue management capacity to reconstruct historical data series, between two connection sessions, in case of null or partial transmission. This operation mode requires the presence of a software architecture that operates as a buffer, acting as an intermediary or as middleware (LEWARnet) between the data consumer (DCPC) and the data producers (sensors and sub-networks of sensors).
The developed middleware also monitors the processes of transmission and data acquisition, recognizing the activity status of the sensors and that of the DCPC, and integrating encryption and data compression functions.
A detailed description of LEWIS can be found in Costanzo et al. (2015, 2016).
The management of information flows, the telematic architecture and the services for data management are entrusted to the DCPC.
The DCPC has been designed and performed according to a complex hardware and software system able to ensure the reliability and continuity of the service, providing advance information of possible dangerous situations that may occur.
In the research project, the DCPC has to ensure the continuous exchange of information among monitoring networks, mathematical models and the command and control center (CCC) that is responsible for emergency management and decision making.
Data flow from the monitoring network was managed according to a communication protocol implemented by the DCPC and named AqSERV. AqSERV was designed considering the heterogeneity of devices of monitoring and transmission networks (single-point and area sensors) and the available hardware resources (microcontrollers and/or industrial computers). AqSERV was devised to link the DCPC database (named LEWISDB) to the monitoring networks, after validation of the authenticity of the node that connects to the center. Data acquisition, before the storage in the database, is validated both syntactically and according to the information content. The procedures for extraction of the information content and validation have been realized differently for single-point and area sensors: the latter require a more complex validation, as they work in a 2-D domain.
The complete management of the monitoring networks by DCPC has been realized through specific remote commands, sent to individual devices via AqSERV, to reconfigure the acquisition intervals or to activate any sensor, depending on the natural phenomena occurring in real time.
The configuration of monitoring networks – composed of devices and sensors, of communication protocol used by each network, and of rules for extraction and validation of information content – is carried out through a Web application that allows for the management of the entire system by the users.
The real-time search for acquisitions is carried out through a WebGIS that has been specifically designed for WSNs but that can be easily extended to classic monitoring networks.
The WebGIS was designed according to the traditional Web architecture,
client–server, by using network services which are Web mapping oriented:
Web server for static data, Web server for dynamic data, server for maps, database for the management of map data.
The test site chosen to experiment the integrated system is located in southern Italy, along a section of the A16 highway, an important communication route that connects Naples to Bari (Fig. 2). The selected section of the A16 highway runs in the SW–NE direction, along the southern Italian Apennines, in correspondence with the valley of the Calaggio Creek, between the towns of Lacedonia (Campania region) and Candela (Puglia region).
Landslides detected through field survey along the monitored section of A16 highway.
The area is tectonically active, but the landscape, characterized by gentle slopes, is mostly influenced by lithologic factors rather than by tectonics. The lithologies outcropping in this area are Pliocene–Quaternary clay, clayey marlstones and more recent (Holocene) terraced alluvial sediments (from clay to gravel). The landslides shown in Fig. 2 are all located in clay or clayey marlstones.
The highway runs on the right flank of the Calaggio Creek at an altitude between 300 and 400 m a.s.l.; the section of interest represents an element at risk in the computation of landslide risk assessment, due to the presence of unstable areas which can potentially affect the communication route (Fig. 2). These unstable areas mainly involve clayey superficial layers.
On 1 July 2014, the GB-InSAR system was installed on the test site. The location of the installation point was selected, taking into account the view of the unstable area and the distance from the power supply network. A covered structure was built to protect the system from atmospheric agents and possible acts of vandalism, in the perspective of a long-term monitoring.
The transmission network was provided by a GSM modem, exploiting the 3G network. In addition to the PC integrated into the GB-InSAR power base, a further external PC was exclusively employed for data after elaboration and transmission.
The system acquired data from the beginning of July 2014 until the end of July 2015.
The installation location allowed the system to detect an area between 40 and
400 m from the its position in the range direction and about 360 m wide in
the azimuth direction. These values, coupled with a 40
The most relevant matter of this monitoring was not as much related to the detection of landslide movements threatening the highway as to how long-term monitoring performed with an instrument providing huge amounts of data could have been run without resorting to large hard drives or to fast Internet connections. In fact, the monitoring area was covered by a 3G mobile telecommunication networks with a limit of 2-gigabyte data transfer per month, and there was the need to reduce the massive data flow produced by the radar.
For this reason, an appropriate data management system (Fig. 3) was developed and is described herein.
Diagram showing the complete data flow from acquisition to final visualization.
The GB-InSAR employed produced a single radar image, consisting of a
This amount of data represented an issue for both store capacity and data transmission.
After being acquired, data were then transferred through LAN connection to the external PC implementing a dedicated Matlab script locally performing the actions described as follows.
In order to reduce the noise normally affecting radar data (especially in vegetated areas), the images acquired every 5 min were also averaged using all data of the previous 8 and 24 h. Then images averaged over 24 h were used to calculate daily displacement maps every 8 h to create 8 h displacement maps, and non-averaged images were used to calculate 5 min displacement maps. These time frames were selected based on the characteristics of the slope movements and signal / noise ratio in the investigated area.
Averaging is also a way to make good use of a high data frequency, since it enables the memory occupied in the database to be reduced as an alternative to their direct elimination.
Each radar image can be represented as in Eq. (1):
One of the major advantages of GB-InSAR is the capability to achieve sub-millimeter precision. However, this can be severely hampered by the variations of air temperature and humidity, especially when long distances are involved. Usually, atmospheric correction is performed by choosing one area considered stable, taking into account that every displacement value different from 0 is due to atmospheric noise and assuming that this offset is a linear function of the distance. Based on this relation the entire displacement map is corrected. In our case the whole scenario has been selected, and then only the potentially unstable zones and those with a weak or incoherent backscattered signal were removed. The remaining areas were then considered stable and therefore were used for calculating the atmospheric effects. This results in a larger correction region that enables a statistical correlation between the atmospheric effects and the distance and therefore the calculation of a site-specific regression function that may not necessarily be linear (Fig. 4).
The color bar is expressed in millimeters; green indicates stable pixels,
while blue and red respectively indicate movement toward and away from the GB-InSAR.
The availability to detect only the line-of-sight (LOS) component of the displacement vector represents one of the main limitations of the GB-InSAR technique. A method to partially overcome this limitation has been applied in this paper, following the procedure described in Colesanti and Wasowski (2006) and later in Bardi et al. (2014, 2016). Other methods have been employed by Cascini et al. (2010, 2013).
Assuming the downslope direction as the most probable displacement path, radar data have been projected in this direction. Input data as the angular values of the aspect and slope have been derived from the digital terrain model (DTM) of the investigated area; furthermore, azimuth angle and incidence angle of the radar LOS have been obtained.
After calculating the direction cosines of LOS and slope (respectively
functions of azimuth and incidence angles and aspect and slope angles) in the
directions of zenith (
The real displacement (
Assuming that the studied landslide actually moves in the downslope direction, the GB-InSAR detectable real displacement percentage ranges between 22 and 60 % (Fig. 5a).
In Fig. 5b, an example of a slope displacement map is shown. Here, cumulated displacement data related to a period between 1 July and 1 November 2014 have been projected in the downslope direction. Data show that the area can be considered stable in the referred period; maximum displacement values of 4 mm in 4 months (eastern portion of observed scenario) can be still considered in the range of stability.
In order to allow for a fast data transfer–velocity threshold comparison, some representative control points were selected, aimed at providing cumulated displacement time series. Control points were retrieved from the same displacement maps calculated as described in Sect. 5.2.2 and therefore can be relative to a time frame of 5 min, 8 h or 24 h.
In the case of noisy data, instead of having a time series relative to a single pixel, these can be retrieved from a spatial average obtained from a small area consisting of few pixels.
Typically, the field of view of a GB-InSAR is larger than the actual area to be monitored. In fact, a portion of the radar image may be relative to the ground, sky and areas geometrically shadowed or covered by dense vegetation. These may be of no interest or even contain no information at all. For the case here studied around 50 % of a radar image had a low coherence and was for all practical purposes unusable. Therefore, a cropping of the ASCII displacement map occurred in order to frame only the relevant area.
The interferometric data generated by GB-InSAR, after the pre-processing and proper correction previously described, are ready for transfer to the DCPC. The transmission of these data to the DCPC is mediated by the middleware, which interrogates the GB-InSAR in order to track the state, detects the newest data, and reorders and marks them to properly build data time series to be transferred to DCPC.
Subsequently, the middleware manages communications with the DCPC, according to the implemented ad hoc protocol. This ensures the security of data providers through encrypted authentication mechanisms; it allows for recovering missing or partially transmitted data, thus avoiding information loss; and it provides data acquired by the sensors to the DCPC in a standardized format, JSON, able to guarantee uniformity between the various information provided by the various sensors types. All these particular features fully justify the adoption of an ad hoc protocol for data transfer, instead of using a standard protocol such as FTP.
The data files produced by the GB-InSAR have already been locally
pre-processed and result in a matrix expressed in ASCII code; the dimensions
of the matrix are known and range from
Eventually the DCPC is entrusted with cumulating the displacements relative to the control points, which are compared with the respective thresholds, and with visualizing the displacement maps as WebGIS layers, thus enabling data validation and the evaluation of the extension of moving surface.
The GB-InSAR is part of a larger early warning system (LEWIS) which also includes other monitoring systems and simulation models. Therefore, to understand how GB-InSAR data can be used in an early warning context, it is necessary to make reference to LEWIS as a whole.
Any information coming from the investigated sites and subsequently processed also by using the simulation models is used to define an intervention model. This is based on the following elements: event scenarios, risk scenarios, levels of criticality and levels of alert.
Event scenarios describe the properties of expected phenomena in terms of
dimension, velocity, involved material and occurrence probability. Occurrence
probability depends on the associated time horizon, which should be equal to
a few hours at most, in the case of early warning systems. Evaluation of
occurrence probability is carried out by using information from monitoring
systems and/or from outputs of adopted mathematical models for nowcasting.
All the properties to be analyzed for event scenarios are listed below; a
subdivision in classes is adopted for each one:
landslide velocity (five classes from slow to extremely rapid); landslide surface (five classes from very small to very large); landslide scarp (five classes from very small to very large); landslide volume (five classes from extremely small to large); thickness (five classes from very shallow to very deep); magnitude (three classes: low, moderate, high), which combines the previous information; involved material (mud, debris, earth, rock, mixture of components); occurrence probability (zero, low, moderate, high, very high, equal to 1).
While some of the aforementioned parameters are determined by geological
surveys, landslide velocity is directly derived from monitoring data (such as
those collected by GB-InSAR). Landslide surface can be determined by
geomorphological observation but is precisely quantified by GB-InSAR, thanks
to its capability to produce 2-D displacement maps.
Risk scenarios can be firstly grouped in the following three classes:
mud and/or debris movements which could induce a friction reduction
between the vehicles and the tar and therefore facilitate slips; road subsidence induced by landslides that could drag or drop vehicles; falls of significant volumes and/or boulders that could crush or cover
vehicles and constitute an obstacle for other vehicles.
For each previous risk scenario, six sub-scenarios can be identified based on
the number of potentially involved infrastructures, carriageways and lanes
(a: hydraulic infrastructures
and/or barriers of either carriageway; b: only the emergency lane of either carriageway;
c: the emergency lane and up to the regular lane of either carriageway;
d: up to the fast lane of either carriageway; e: up to the fast lane of the opposite carriageway;
f: up to the regular lane of the opposite carriageway).
Thus, the total number of possible risk scenarios is 18 (Fig. 6), indicated
with a couple of letters (capital and small).
Top and middle: possible risk scenarios involving scenario A (landslides that could reduce friction) to increasing sectors of the highway. Bottom: combinations of scenarios with several types of phenomena that affect the emergency lane, regular lane and fast lane.
The following information is provided to DCPC:
measurements from sensors, model outputs, state 0: no variation, state 1: small variation, state 2: moderate variation, state 3: high variation.
and four states are identified for each of them:
In practice, for the GB-InSAR, such states are delimited by fixed velocity
values (thresholds). In this application values have been selected according
to the gathered data, the first threshold being just above the instrumental
noise; the remaining have been set based on expert judgement waiting for a
more robust calibration, which is possible only after at least a partial
mobilization of the slope. Anyway, the system is open to any method for
determining thresholds (Crosta and Agliardi, 2003; Du et al., 2013;
Carlà et al., 2016a) and also to the use of other parameters
(acceleration for example).
Besides information from sensors and models, other information is obtained from meteorological and hydrological models (named “indicators”).
Indicators comprise weather forecasting and output of the Forecasting of Landslides Induced by Rainfall (FLaIR) and Saturated Unsaturated Simulation for Hillslope Instability (SUSHI) models (Sirangelo et al., 2003; Capparelli and Versace, 2011) on the basis of observed and predicted (for the successive 6 h) rainfall heights.
Two states are defined for indicators:
state 0: no variation or not significant, state 1: significant variation. state (0, 1) of indicators (INDs), state (0, 1, 2, 3) of sensors and models running for the specific highway
section (SEN),
To sum up, DCPC has the following information at any moment:
and, on the basis of these states, four different decisions can be made by
DCPC, one of which with three options.
All the possible decisions are illustrated in Table 1, in which the weight of the several sensors is assumed to be the same. Based on the notices of criticality levels provided by the DCPC, and on its own independent evaluations, the CCC issues the appropriate warning notices (surveillance, alert, alarm and warning) and makes decisions about the consequent actions.
DCPC possible decisions.
The information of each sensor and the results produced by the models are used to assess, in each instant, the occurrence probability of an event scenario in the monitored areas and the possible risk scenarios.
This combination of heterogeneous data was carried out by identifying for each sensor and model typical information (displacement, precipitation, inclination etc.); evaluating the state at each instant, according to a threshold system; and combining this result for all sensors placed in a monitored geomorphological area.
The result is constituted by the occurrence probability of an event scenario, which is associated with a specific action by the DCPC. In particular, if the occurrence probability is low, moderate or high, it is necessary to issue a notice of criticality (ordinary – level 1; moderate – level 2; high – level 3) to the CCC.
The DCPC sends two types of information:
criticality state of the single monitored geomorphological unit, criticality state of the whole area.
The adopted communication protocol between the two centers for the exchange
of information was carried out through a Web service provided by the CCC,
using the classes and attributes of the methodology named DATEX II (which is
a protocol for the exchange of traffic data). The use of the Web service
allowed ensuring the interoperability of data between the two centers,
regardless of the hardware and software architecture used, through a
persistent service capable of ensuring an immediate restoration of the
connections, in case of malfunction, and a continuous monitoring between the
two centers, even in the absence of criticality.
The GB-InSAR is a monitoring tool that is becoming more and more used in landslide monitoring and early warning, especially thanks to its capability to produce real-time, 2-D displacement maps. On the other hand, it still suffers from some drawbacks, such as the limitation of measuring only the LOS component of a target's movement and logistic issues like those owing to a massive production of data that may cause trouble for both storing capacity and data transfer. In particular, the latter is a more and more common problem of advanced technologies that are able to produce high-quality data with a high acquisition frequency, which may leave the problem of finding the balancing between exploiting all the information and at the same time avoiding unnecessary redundancy.
These problems were addressed when a GB-InSAR was integrated within a complex early warning system (LEWIS) and only a limited Internet connection was available. This situation required that a series of pre-elaboration processes and data management procedures take place in situ in order to produce standardized and reduced files, carrying only the information needed when it was needed. The procedures mainly concerned the transmission of data averaged over determined time frames, proportionate with the kinematics of the monitored phenomenon. Previously, transmission data were also corrected (in terms of both atmospheric noise and LOS) and reduced, by filtering out the information relative to the amplitude of the targets, by eliminating the areas not relevant for the monitoring and by transforming the matrices into strings.
As a result, GB-InSAR data converged into the early warning system and contributed to it by producing displacement time series of representative control points to be compared with fixed thresholds. Displacement maps were also available for data validation by expert operators and for retrieving information relative to the surface of the moving areas.
The present paper concerns a methodology. Some of the data used in this paper are no longer available; the remaining available data can be requested from the corresponding author.
The authors declare that they have no conflict of interest.
This article is part of the special issue “Landslide early warning systems: monitoring systems, rainfall thresholds, warning models, performance evaluation and risk perception”. It is not associated with a conference.
This research is part of the project “LEWIS (Landslides Early Warning Integrated System): An Integrated System for Landslide Monitoring, Early Warning and Risk Mitigation along Lifelines”, financed by the Italian Ministry of Education, Universities and Research and co-funded by the European Regional Development Fund, in the framework of the National Operational Programme 2007–13 “Research and Competitiveness”, grant agreement no. PON01_01503.The authors are thankful to Giuseppe Della Porta and his colleagues from Autostrade S.p.A. for their availability in permitting and supporting the installation and maintenance of the GB-InSAR along the A16 highway. Edited by: Stefano Luigi Gariano Reviewed by: two anonymous referees