Baseline monitoring of volcanic regions with little recent activity: application of Sentinel-1 InSAR to Turkish volcanoes

Abstract

Volcanoes have dormancy periods that may last decades to centuries meaning that eruptions at volcanoes with no historical records of eruptions are common. Baseline monitoring to detect the early stages of reawakening is therefore important even in regions with little recent volcanic activity. Satellite techniques, such as InSAR, are ideally suited for routinely surveying large and inaccessible regions, but the large datasets typically require expert interpretation. Here we focus on Turkey where there are 10 Holocene volcanic systems, but no eruptions since 1855 and consequently little ground-based monitoring. We analyse data from the first five years of the European Space Agency Sentinel-1 mission which collects data over Turkey every 6 days on both ascending and descending passes. The high relief edifices of Turkey’s volcanoes cause two challenges: 1) snow cover during the winter months causes a loss of coherence and 2) topographically-correlated atmospheric artefacts could be misinterpreted as deformation. We propose mitigation strategies for both. The raw time series at Hasan Dag volcano shows uplift of ~ 10 cm between September 2017 and July 2018, but atmospheric corrections based on global weather models demonstrate that this is an artefact and reduce the scatter in the data to < 1 cm. We develop two image classification schemes for dealing with the large datasets: one is an easy to follow flowchart designed for non-specialist monitoring staff, and the other is an automated flagging system using a deep learning approach. We apply the deep learning scheme to a dataset of ~ 5000 images over the 10 Turkish volcanoes and find 4 possible signals, all of which are false positives. We conclude that there has been no cm-scale volcano deformation in Turkey in 2015–2020, but further analysis would be required to rule out slower rates of deformation (< 1 cm/yr). This study has demonstrated that InSAR techniques can be used for baseline monitoring in regions with few historical eruptions or little reported deformation.

Introduction

Many volcanoes have periods of dormancy of decades to centuries between eruptions and eruptions from volcanoes with little historical record of activity are highly problematic because the populations and emergency managers have no experience of volcanism. Recent examples include the eruptions of Chaitén (Chile) in 2008, Sinabung (Indonesia) in 2010 and Soufrière Hills (Montserrat) in 1995. The Global Volcanism Program currently list 10 volcanic systems in Turkey with activity in the Holocene (Table 1), including 3 with historically recorded eruptions (GVP 2013). However, there have been no eruptions since 1855 and consequently there has been little investment in ground-based monitoring systems despite their proximity to large population centres in central Anatolia. The United Nations Global Assessment of Risk ranked Turkey at 14th for overall volcanic threat out of 95 volcanically active countries based on its historical record of volcanic eruptions and large population exposure. Thus, baseline monitoring of Turkish volcanoes remains important, despite the lack of eruptions in the last century.

Table 1 Table of Turkish Volcanoes (from West to East) with Holocene activity (GVP 2013) and the corresponding frame numbers for ascending and descending passes available from the COMET LiCSAR processing system https://comet.nerc.ac.uk/COMET-LiCS-portal/. Note, volcano names are given according to current Global Volcanism Program Database. The locally approved names and spelling are given in parentheses and used throughout this document

Satellite-based remote sensing is increasingly used as a global tool for monitoring volcanoes (Furtney et al. 2018; Anantrasirichai et al. 2018). Satellite systems are particularly suited to providing baseline information for large regions (e.g. Biggs et al. 2011; Pritchard and Simons 2004) or providing information during eruptions when ground-based equipment cannot be deployed or repaired (Biggs et al. 2010; Sigmundsson et al. 2010; Arnold et al. 2019). Satellite systems are often the only available source of monitoring data when 1) ground based access is limited, for example, in areas of conflict or remote regions, or 2) resources are limited, for example in developing countries, or 3) where there have been few historical eruptions and political awareness of volcanic risks is low. ESA’s Sentinel-1 mission freely distributes radar images over Turkey every six days and this could form the backbone of a baseline monitoring system for detecting deformation.

In this article, we report preliminary findings from the first five years of the Sentinel-1 mission, which have been automatically processed by the LiCSAR system (Morishita et al. 2020). In particular, we demonstrate that for the high relief edifices of Turkish volcanoes, it is necessary to account for the effects of atmospheric water vapour and seasonal snow cover when identifying and measuring deformation. Finally, we consider how the dataset could be used operationally and in real time and develop two strategies for dealing with the large volumes of data produced. First, we design an easy to follow image classification scheme designed for non-specialist monitoring staff, and second we apply an automated flagging system using a deep learning approach.

Background: volcanic activity in Turkey

According to the Global Volcanism Programme, there have been 37 eruptions recorded in Turkey during the last 10,000 years (GVP 2013) (Fig. 1). Three volcanoes have erupted historically (GVP 2013): Nemrut Dağı in 1441, 1597, 1692 and 1881 with recent activity characterised by the formation of lava domes, basaltic and rhyolitic lavas (Karakhanian et al. 2002; Ulusoy et al. 2019a; Aydar et al. 2003). A gas and ash eruption occurred at Tendürek Dağı in 1855 (Karakhanian et al. 2002) and a VEI 3 eruption occurred at Ağrı (also known as Ararat) in 1840, and an uncertain eruption in 1783. Turkey is also susceptible to volcanic hazards including ash and volcanogenic tsunamis from neighbouring countries with volcanoes (Greece, Syria, Iran, Azerbaijan, Armenia and Georgia).

Fig. 1

Location of volcanoes and major cities in Turkey. The 200 km buffer zone indicates areas in neighbouring countries that could potentially be affected by eruptions from Turkish volcanoes. 1. Kula Volcanic Field; 2. Gölcük; 3. Karapınar Volcanic Field; 4. Hasan Dağı; 5. Göllüdağ; 6. Acıgöl; 7. Erciyes Dağı; 8. Karacadağ; 9. Nemrut Dağı; 10. Süphan Dağı; 11. Girekol Tepe; 12. Tendürek Dağı; 13. Ağrı; 14. Kars Plateau

Geological studies suggest numerous Holocene eruptions have occurred in Turkey. In eastern Anatolia, tephras in Lake Van (Sumita and Schmincke 2013b; Sumita and Schmincke 2013a) together with historic events indicate 27 eruptions of Nemrut Dağı (GVP 2013), making this Turkey’s most frequently active volcano. In addition to the 1855 eruption of Tendürek Dağı, radiocarbon dating shows an eruption also occurred there in 550 BC at (GVP 2013). Ararat (Ağrı) has eruptions recorded at 2500–2400 BC and in the 3th–fourth century (Karakhanian et al. 2002). In central Anatolia, Acıgöl, near the city of Nevşehir, has six possible eruptions, most recently in Roman times. Both Erciyes Dağı, near the populous city of Kayseri, and Hasandağ have several possible Holocene eruptions (Atıcı et al. 2018; Sarıkaya et al. 2019; Schmitt et al. 2014) and potential further events are suggested by archaeological evidence and a widespread Holocene tephra that can be traced as far as Lebanon (Hamann et al., 2010). Kula volcanic field located in western Anatolia erupted during the Bronze Age (Ulusoy et al. 2019b).

The absence of eruptions for over a century could suggest the volcanic risk is not an issue in Turkey. However, the historic record includes at least one eruption in Turkey for each of the centuries between 1100 and 1900 AD, suggesting that the last 170 years of quiescence is unusually long. Combined with the rapid population growth since the last eruption in 1855, baseline monitoring is important for all Turkey’s volcanoes, despite the lack of recent activity.

By combining data from a range of satellites, it is possible to construct an archive dating back to the 1990’s and this provides the opportunity to conduct a 20 year baseline deformation survey identifying periods of unrest (e.g. Pritchard and Simons 2004; Reath et al. 2019b; Biggs et al. 2011). Surveys such as these have increased the number of volcanoes known to be deforming five-fold in the last 20 years (Biggs and Pritchard 2017). The global compilation of Furtney et al. (2018) reports 306 volcanoes with satellite-detections: 147 of deformation, 146 with thermal anomalies and 173 with SO₂ degassing. However, in Turkey, there are no records of satellite-detected degassing or thermal anomalies, and only one example of deformation. At Tendürek, Bathke et al. (2015, 2013) report subsidence at a rate of 1 cm/yr attributed to contraction of a sill at a depth of 4.5 km below the summit causing ‘sliding trapdoor’ motion along the caldera ring faults.

Methods: satellite data and processing

There are four types of satellite observations that are particularly relevant for volcano monitoring: 1) Deformation measurements made from radar imagery using a technique called Interferometric Synthetic Aperture Radar (InSAR) (e.g. Biggs et al. 2014; Pritchard et al. 2018; Richter and Froger 2020) 2) Measurements of active or passive degassing using either infrared or UV spectrometers (e.g. Carn et al. 2016); 3) Measurements of thermal anomalies associated with erupted material (e.g. Wright et al. 2004; Coppola et al. 2016) or fumarole activity (Reath et al. 2019a; Jay et al. 2013), 4) Mapping of volcanic products using visual observation of footprints, or topographic measurements for height changes (Poland 2014; Arnold et al. 2019; Meyer et al. 2015). Furtney et al. (2018) showed that satellite-based deformation measurements are dominantly pre-eruptive, while degassing and thermal measurements are dominantly co-eruptive. Since there have been no eruptions in Turkey during the satellite era, we focus on deformation measurements using InSAR.

The InSAR method measures the phase change between two Synthetic Aperture Radar images to produce maps known as interferograms. The phase change of the interferogram, ∆ϕ_i, is composed of several components, primarily deformation of the ground surface, ∆ϕ_def, changes in the viewing geometry, ∆ϕ_geom and delays associated with the radar path through the atmosphere, ∆ϕ_atm. The changes in viewing geometry can be corrected using accurate orbit information and a digital elevation model.

Here we use data from the Sentinel-1 constellation which carries C-band radar that operates in TOPSAR mode. Sentinel-1A was launched in 2014 and provides images over Turkey every 12 days. Since the launch of Sentinel-1B in 2017, the revisit time has reduced to 6 days. The data is freely available from the European Space Agency, automatically processed by the LiCSAR system described in Lazecky et al. (2020) and made available online at https://comet.nerc.ac.uk/COMET-LiCS-portal/. The Turkish Volcanoes can be covered using 7 frames of ascending data and 7 frames of descending data (Table 1). We use a least-squares approach to retrieve displacement maps at each date of acquisition (Schmidt and Burgmann 2003).

Despite the wealth of data, there are several challenges associated with using InSAR data for volcano monitoring. Atmospheric effects, primarily due to water vapour in the troposphere cause anomalies which can obscure deformation, or even lead to misinterpretation of the signals (Yip et al. 2019; Beauducel et al. 2000). These effects are particularly challenging at high relief, tropical volcanoes where the stratification of water vapour can lead to artefacts that correlate to topography (Ebmeier et al. 2013a; D'Oreye et al. 2008). Even though Turkey lies in an arid region, several of the volcanoes are very high relief and the stable continental climate may lead to persistent artefacts.

The second challenge is changes in the surface properties which leads to poor coherence between images. In tropical regions, poor coherence is typically caused by dense tropical vegetation, which scatters the C-band radar waves. The use of longer-wavelength radar systems which can penetrate vegetation, such as the ALOS L-band radar, has proved particularly successful under these circumstances (Ebmeier et al. 2013b; Chaussard and Amelung 2012; Fournier et al. 2010). In contrast, poor coherence in polar regions including Alaska and Iceland is typically caused by snow cover, leading some authors to concentrate on summer only acquisitions (e.g. Lu et al. 2010), or the regions surrounding ice caps (e.g. Sigmundsson et al. 2015). Turkey lies in an arid region, but the high volcanic peaks still experience snow fall during winter suggesting the processing strategy may need to be adapted to maintain coherence.

Results: use of InSAR at Turkish volcanoes

In this section, we discuss the suitability of InSAR for providing baseline monitoring for the volcanoes in Turkey, and consider appropriate strategies for data processing. We use Hasandağ in central Turkey as a test case, as it has good data availability but also high elevation (> 3200 m), meaning it is representative of the atmospheric artefacts and snow cover which may prove challenging for routine volcano monitoring using InSAR.

Coherence

Turkey is located in an arid environment and the region is not heavily vegetated, meaning the coherence is generally excellent and InSAR has been used for many tectonic studies (e.g. Cakir et al. 2005; Burgmann et al. 2002; Wright et al. 2001; Weiss 2020). However, the high peaks volcanic peaks are covered with snow during the winter months, reducing coherence.

Figure 2 shows example interferograms from Hasandağ. Figure 2a shows an interferogram constructed using acquisitions in May 2020 when there was no snow cover, whereas Fig. 2b is an interferogram constructed using acquisitions in January 2020, a time of year when snow cover is common. Coherence can be measured as the cross-correlation between the complex values in the SAR image with the theoretical range of values between 0 (no correlation) and 1 (perfect correlation). For the LICSAR processing, coherence thresholds are typically set at 0.15–0.2, with values > 0.5 common in urban and desert environments. Figure 2c shows the coherence image for the summer acquisitions, with values > 0.8, and the corresponding interferogram (Fig. 2a) shows continuous data coverage. In contrast the coherence image for the winter acquisitions (Fig. 2d) shows values < 0.1 over the volcanic edifice indicating snow cover on one or both acquisitions. The corresponding areas of the interferogram (Fig. 2b) have random phase values and cannot be used to infer anything about the surface deformation.

Fig. 2

Example interferograms from Hasandağ showing seasonal effects on coherence a) interferogram showing good coherence in the snow-free summer months (6-18th May-2020) b) interferogram showing poor coherence when the peaks are covered in snow during the winter months (7–31 Jan 2020), c) coherence image corresponding to the interferogram shown in a). f) coherence image corresponding to the interferogram shown in b)

Figure 3 shows seasonal variation in coherence at Hasandağ. The mean and standard deviation of the coherence for 389 automatically-generated interferograms between October 2014 and May 2020 with time intervals of 12 and 24 days are plotted according to the date-of-year of the primary acquisition. The dataset contains 7 faulty interferograms which have coherence of 0 and can be easily discarded. A further 15 have average coherence values < 0.2 and these occur in the winter months and are associated with almost complete snow cover and cannot be used. Average values < 0.4 are relatively common and occur 103 times (27% of the dataset). In the winter months (Dec-Feb), these are attributed to partial snow cover over the high peaks, whereas in summer (Apr-Jul), it is due to vegetation growth and agriculture at lower elevations. However, not all the interferograms during these time periods are affected, so rather than skipping these periods entirely (e.g. Lu et al. 2010), it may be more appropriate to process interferograms all year round and discard those affected by poor coherence. In some cases, it may be necessary to process additional interferograms using snow-free acquisitions and longer time-spans to ensure that all time periods are covered. An alternative approach is to reprocess the dataset using a persistent scatterer method (e.g. Hooper et al. 2007; Ferretti et al. 2001), which has proven effective in semi-vegetated areas, but may not be appropriate when deep snow buries all infrastructure.

Fig. 3

Interferometric coherence in a 50 km box around Hasandağ volcano, Turkey a) Mean and standard deviation of the coherence values based on 5 years of observations (2014–2020) plotted as a function of season. Green is interferograms spanning 12 days and red is 24 days

Identification of Atmospheric Artefacts

Atmospheric effects can be caused by variations in ionospheric or tropospheric properties. For C-band data at this latitude, tropospheric effects typically dominate (Liang et al. 2019) and are typically separated into a dry component associated with changes in pressure and temperature and a wet component associated with water vapour (Elliott et al. 2008; Jolivet et al. 2014; Doin et al. 2009). These can be further subdivided into short-wavelength ‘turbulent’ effects, and topographically-correlated effects associated with the stratification of atmospheric water vapour. At high-relief, conical volcanoes, these topographically-correlated delays cause patterns very similar to those expected from inflation or deflation of a magma reservoir (e.g. Ebmeier et al. 2013a; Beauducel et al. 2000).

Figure 4 shows examples of atmospheric artefacts from Hasandağ. The interferograms in Fig. 4a and b show a strong correlation with topography, with anomalies located over high topography. This can also be visualised by plotting the phase of each pixel in an individual interferogram against the topography of that pixel (e.g. Ebmeier et al. 2013a, Beauducel et al. 2000). Figure 4c and d show a spatially smooth pattern of noise, typical of transient water vapour features associated with turbulence in the troposphere and which tend to have wavelengths of < 10 km (e.g. Hanssen 2001; Emardson et al. 2003).

Fig. 4

Example interferograms from Hasandağ showing common atmospheric artefacts a) 19-Aug-2015 to 30-Oct-2015 and b) 19-Nov-2018 to 1 Dec-2018 show stratified tropospheric artefacts that correlate to the steep topography of the volcanic edifice c) 6-Apr-2020 to 24-Apr-2020 and d) 14-Aug-2017 to 20-Aug-2017 show turbulent water vapour artefacts. All images are from ascending track 87

In previous studies, a pair-wise logic test has be useful in identifying acquisitions with particularly strong atmospheric artefacts (e.g. Massonnet and Feigl 1995; Ebmeier et al. 2013a). If the signal is associated with a strong atmospheric anomaly on a particular day the sign will change between successive interferograms. For example, we construct interferograms between images A, B and C when there is a strong atmospheric anomaly on acquisition date B. A negative anomaly appears on the volcanoes in interferogram B-A, whereas this anomaly appears positive in interferogram C-B. We expect both turbulent and stratified atmospheric anomalies to follow this pattern, but true deformation signals would reflect a permanent phase shift and the sign would not reverse in successive interferograms. However, we find that the pairwise logic approach performs poorly at identifying atmospheric artefacts in the data at Hasandağ. We attribute this to the stable continental climate where the atmospheric artefacts are low-level and persistent rather than extreme, infrequent events associated with, for example, the passage of a storm.

Correction of Atmospheric Artefacts

Several methods have been proposed to correct for atmospheric errors, and these can be divided into three categories: 1) empirical corrections based on the correlation between phase and topography (e.g. Bekaert et al. 2015; Beauducel et al. 2000), 2) corrections based on external datasets such as GPS or MERIS (e.g. Li et al. 2005; Li et al. 2009) and 3) corrections based on global or regional weather models (e.g. Yu et al. 2018; Jolivet et al. 2014; Parker et al. 2015; Doin et al. 2009). However, each of these has advantages and disadvantages when applied to individual volcanoes (e.g. Stephens et al. 2020), and the priority for volcano monitoring is confidently identifying which signals are atmospheric artefacts, and which are really deformation.

Here we follow the approach of Albino et al. (2020) and test two types of atmospheric corrections: 1) the phase elevation relationship and 2) GACOS corrections (Yu et al. 2018), which are based on the ECMWF High Resolution Weather Model (HRES- ECMWF). We find that while GACOS corrections work well on some example interferograms (Fig. 5a-c) and the phase-elevation corrections works well in others (Fig. 5d-f), neither method works well for every case. Having analysed a total of 288 interferograms, we find that the best approach to the entire dataset is to 1) apply the GACOS correction, 2) remove any long-wavelength ramps and 3) apply a phase-elevation correction. This gives a quality improvement, Q, of 22%, where Q is defined by the difference in standard deviation before and after correction, normalised by the standard deviation before correction (Albino et al. 2020).

Fig. 5

Examples of atmospheric corrections to individual interferograms from track 167 which covers Hasandağ, Acıgöl-Nevsehir and Göllüdağ and encompasses elevations from 600 m–3600 m. a-c) example where b) the GACOS correction works better than c) the phase-elevation correction; d-f) example where e) the GACOS correction works worse than f) the phase-elevation correction

While these methods may be effective at reducing the effects of stratified water vapour fields and, in the case of weather models, long-wavelength atmospheric anomalies, they do not have the resolution to mitigate short-wavelength anomalies such as those associated with turbulence in the troposphere. However, since such anomalies are transient, we can assume that there is no correlation in time and they can be reduced stochastically. To illustrate, if we sum N independent interferograms of equal length, the signal will increase by a factor of N, whereas the noise will only increase by a factor of √N. The overall result is an increase in the signal to noise ratio by a factor of √N. In practice, interferograms are constructed using common acquisitions, meaning that the interferograms are not independent and chain stacking or time series approaches are more appropriate (Morishita et al. 2020; Berardino et al. 2002; Hooper et al. 2012).

Figure 6 shows the time series with and without atmospheric corrections for Hasandağ. The uncorrected time series has a seasonal component consistent with stratified atmospheric artefacts which could possibly show a possible uplift of ~ 10 cm starting in October 2017 (Fig. 6a). In comparison, the corrected time-series is much cleaner, and show no seasonal or other patterns (Fig. 6b). The standard deviation is reduced from 2.5 cm prior to correction to < 1 cm afterwards. We conclude that there is no volcanic deformation signal at Hasandağ during the period 2016–2018.

Fig. 6

Examples of a) uncorrected and b) corrected Sentinel-1 InSAR timeseries from Hasandağ (34.17E 38.13 N) on the descending pass during the period 2016–2018. The point F is located away from the volcanic edifice (34.33E 38.39 N) and serves as a reference point for comparison

Discussion: towards operational monitoring

Many volcano observatories do not employ a specialist in InSAR data analysis, particularly in countries where there have been few historical eruptions and monitoring is conducted at a baseline level only. Lack of expertise is often cited as a barrier to the uptake of satellite technology and several capacity building efforts are ongoing to address this issue. An alternative is to develop automated systems that require little user input (Meyer et al. 2015; Lazecky et al. 2020). Here we consider two approaches that are particularly suited to regions with little record of volcano deformation who wish to use satellite data for baseline monitoring (i.e. flagging unusual activity) rather than modelling and forecasting deformation signals. In this case, our approach is based on image classification, either using a simplified scheme designed for non-specialists, or automated systems based on machine learning we then consider how the existing ground-based monitoring, (including the national seismic and GNSS networks) could be augmented if the satellite system were to detect unrest.

Guide to interpretation for non-specialists

The aim of this section is to provide a simple and easy-to-follow classification scheme by which non-specialist monitoring staff can interpret images produced by automated systems such as LiCSAR. We adapt the flowchart of Ebmeier et al. (2013a) which was designed based on an L-band radar survey of Central America (Fig. 7). The flow chart includes decision points (shown as diamonds), causes of signals (white boxes), example data from Hasandağ, and mitigation strategies (pink boxes). In many cases, the mitigation strategies may be only be feasible retrospectively, or require specialist techniques, in which case, the images could be discarded from any real-time analysis.

Fig. 7

Flowchart illustrating the process of analysing automatically generated interferograms, including checks and corrections for low coherence, topographic correlations and low signal-to-noise ratios before identifying possible signals. a-c Examples of common artefacts illustrated using examples from Hasandağ volcano, Turkey. a low coherence due to snow cover, b topographic correlation due to stratified water vapour and c low signal-to-noise ratio due to turbulent atmospheric effects

The first step is to identify whether the data is coherent. If the coherence over the summit is low, this suggests the interferogram is affected by snow fall on one or both of the acquisition dates, and needs to be discarded. If the time period covered is particularly critical, longer-duration interferograms which use different acquisitions dates may avoid the snow cover. The next step is to see whether the signal is correlated to topography, either by visual inspection, or by plotting the phase change against elevation. If there is a significant topographic dependence, the interferogram could be affected by a stratified atmospheric artefact, particularly if similar signals are seen over multiple peaks. Two mitigation strategies are available – either using an empirical correction based on the gradient of the phase elevation plot, or using a global weather model from an online service such as GACOS (http://ceg-research.ncl.ac.uk/v2/gacos/). The next stage is to consider the signal-to-noise ratio of the interferogram. In the case of slow deformation signals and short duration interferograms, turbulent atmospheric artefacts may obscure any real deformation. In this case, stacking or time-series approaches could be used retrospectively to improve the signal-to-noise ratio and lower the detection threshold. Having ruled out the other possibilities, any remaining signal is likely to be caused by real deformation. In this case, the data should be compared with any other available information (e.g. from the regional seismic network, other InSAR tracks or satellite systems) and modelled to identify the causative source.

Automatic detection using machine learning

Even with a simplified interpretation framework, the number of satellite images available poses a challenge for data management. The Sentinel-1 satellite has produced ~ 5000 images over the 10 Turkish volcanoes in the last 5 years, a rate of about 20 per week. Machine learning tools are increasingly important in handling such large datasets (Gaddes et al. 2019; Sun et al. 2020; Valade et al. 2019), and deep learning methods such as convolutional neural networks are especially useful for classifying images. Anantrasirichai et al. (2018, 2019a) developed a transfer learning approach to identifying volcano deformation in global datasets of Sentinel-1 imagery. The pre-trained network is re-trained using synthetic images to discriminate between real deformation and atmospheric artefacts. The best performing model considers just two classes: images composed of turbulent and stratified atmospheric components only (S + T) versus images containing deformation signals plus atmospheric artefacts (D + S + T). From a global dataset of > 30,000 images, Anantrasirichai et al. (2019a) identified just 52 positive results, of which 41 were true positives. While not perfect, this reduces the number of images requiring manual interpretation by two orders of magnitude, thus making routine analysis of large datasets tractable.

Here we apply that framework to the ~ 5000 images acquired over Turkish volcanoes in 2015–2020 (Table 2). The images have a range of timespans from 6 to 120 days, with a median value of 18 days. Since the algorithm is trained to distinguish between atmospheric artefacts and real deformation, we input the images prior to atmospheric correction. The algorithm identified just 4 positive results (Fig. 8, Table 2): two from Hasandağ in June 2015 and May–June 2019, one from Erciyes Dağı in October 2019 and one from Karacadağ in Feb-March 2019. Only one of these, the image from Erciyes Dağı had a probability > 0.9, while the images at Karacadağ and Hasandağ in 2019 only had very small areas with probabilities > 0.5 (Fig. 8). Since any real deformation would be expected to span multiple interferograms and these are all isolated images, these can be fairly simply assigned to the false positive category. The false positive rate of < 0.1% is consistent with the results of the global analysis (Anantrasirichai et al. 2019a), and dramatically reduces the need for manual inspection.

Table 2 Deep learning analysis of deformation at Turkish Volcanoes based on the method of Anantrasirichai et al. (2018, 2019a). The Convolutional Neural Network outputs two classes: D + S + T contains a mixture of deformation and atmospheric artefacts while S + T contains only atmospheric artefacts. The images in the D + S + T class are then checked by an expert and assigned to the true or false positive classes. In the case of Turkey, all images were either S + T or false positives. All volcano names are given using locally approved names and spelling with the GVP equivalents given in Table 1

Fig. 8

Four false positive outputs of the deep learning algorithm applied to ~ 5000 interferograms covering the 10 volcanoes in Turkey. The brighter yellow means higher probability. The blue contour represents a 50% probability and the green contours 90%

This shows that none of Turkey’s volcanoes are currently experiencing significant deformation. However, the detection threshold of the simple image-by-image approach is quite high as 3–4 cm in a 6 day interferogram corresponds to a rate of several meters per year. The detection limit can be reduced applying pre-processing techniques including time series analysis and overwrapping, giving a final detection threshold of < 1 cm/yr for a 5 year dataset (Anantrasirichai et al. 2019b). Similarly, the accuracy of the detection algorithm (number of false positive) can be improved by using global weather models to reduce the atmospheric artefacts (Anantrasirichai et al. 2019a).

Integration with other monitoring techniques

The ability of satellite systems to routinely survey large regions is invaluable for baseline monitoring, particularly for regions with few historical eruptions where investment may be limited. However, despite rapid technological development, satellite systems still cannot measure the range of phenomenon or achieve the same temporal and spatial resolution as ground-based sensors and are not expected to replace the need for ground-based networks (Ebmeier et al. 2018). When satellite (or other) systems detect unrest, additional monitoring may need to be deployed. Turkey has a well-developed regional GNSS network (e.g. Weiss 2020; Nocquet 2012), and the recent TurkVolc project set up campaign GNSS sites at Hasandağ and Erciyes Dag which could be rapidly upgraded to continuous sites if unrest were detected. Similarly, Turkey has a well-established seismic network designed for monitoring earthquake activity, with stations within 55 km of each of the volcanoes, and 8 volcanoes with stations within 20 km (Atici et al., pers. comm.). Recently, these have been supplemented by seismometers specifically designed to monitor volcano seismicity at Erciyes Dağı and Nemrut Dağı. Together these systems provide baseline monitoring, which could be rapidly augmented and densified if unrest were detected.

Conclusions

We conclude that the Sentinel-1 InSAR dataset has the potential to be used for baseline monitoring of Turkish volcanoes. The satellites acquire radar images every 6 days on both ascending and descending passes and the raw data is downloadable in near real time from the Copernicus Open Access Hub (https://scihub.copernicus.eu/) and processed products from the LiCSAR processing system (https://comet.nerc.ac.uk/comet-lics-portal/).

Our analysis of the first 5 years of Sentinel-1 data (2014–2020) has identified two primary challenges for the use of InSAR data for baseline monitoring of the high elevation Turkish volcanoes 1) seasonal snow cover over the high peaks which reduces interferometric coherence and 2) topographically-corrected artefacts associated with stratified tropospheric water vapour which can be misinterpreted as volcano deformation. We propose mitigation strategies for these issues including adaptation to the processing scheme, the use of atmospheric corrections based on global weather models and multi-interferogram approaches such as time series and stacking. We apply these to a retrospective analysis of 3 years of data from Hasandağ volcano and find the uncorrected time series shows a seasonal pattern with up to 10 cm of apparent uplift. In contrast, the corrected time series shows no systematic deformation and variations with a standard deviation of < 1 cm.

Since these approaches may not be feasible in real-time, we demonstrate two possible approaches for operation monitoring based on classification of individual images, either by non-specialist monitoring staff, or by deep learning algorithms. At present, the detection threshold of these methods is quite high, particularly for short-duration interferograms, but the field of machine learning is developing rapidly and techniques to improve the detection accuracy and threshold are likely to be available soon.

In conclusion, our analysis of the first 5 years of data from the Sentinel-1 mission finds no cm-scale deformation at any of the Turkish volcanoes, and we propose processing and analysis strategies adapted for using InSAR data for baseline volcano monitoring in Turkey.