Thus, in this paper, the interferometric synthetic aperture radar (InSAR) data is being used which comprise of an interferogram generated from two SAR images taken at a different date and is compared with its associated coherence image to remove incoherent regions such as vegetation from the interferogram. Here, the images used are taken from the Sentinel-1 radar. This paper introduces a technique that is low-cost compared to other techniques available such as Persistent Scatterer Interferometry (PSI) for detecting the volcano deformation using the InSAR data and computes a score value to determine whether the deformation has taken place or not at that particular time. The algorithm is being tested on three recently active volcanoes Erta Ale, Piton de la Fournaise and Kilauea which triggered lava flows and finally, the results of these three volcanoes are compared by generating a time series for the score values for each volcano. In earlier times monitoring and investigating volcanos required intensive fieldwork (Pinel et al., 2014). During volcanic eruptions and unrest, frequent observation of the volcano is required to monitor the change in activity, in order to update hazard assessments and protect local populations. Thus up to date observations of the surface is crucial for current hazard level (Vassileva et al., 2017). Observations of the volcano can be made using a variety of techniques and sensors, but most of the methods have limitations. For e.g., Visual observations do not provide accurate results during the night and poor weather conditions. Most of the current ground-based techniques have limitations too as they work for a single location and are placed at a distance from active vents. SAR observations thus complement other monitoring techniques (Arnold et al.,2018). SAR is a radar imaging technique that uses microwaves to produce high-resolution images and works on the Doppler-Effect. It works in all weather conditions as it is not dependent on natural illumination and has low atmospheric absorption property. It also captures ten to hundred kilometers of wide areas repeatedly from days to weeks and is, therefore, most suited for volcanic systems. As a single SAR image is not very useful a technique called InSAR is used for monitoring volcanoes.
InSAR is a technique that uses two SAR images taken at different time of the same area and produces an interferogram that represents the phase difference of two SAR images. Despite, various methodology available that uses the InSAR (see Section 2) all these methods are complicated and requires a large dataset for better performance. This paper introduces a low-cost technique for detecting surface deformation due to volcanic activities. This method has an advantage over other methods as it is easy to implement and works well even when the dataset is small. The data provided by Sentinel-1 radar for three main volcanoes Erta Ale at Ethiopia, Kilauea at Hawaii and Piton de la Fournaise at Ile de La Runion for the time period October 2016- June 2018 are taken into account. Sentinel-1 has an advantage over the other radars due to its ability to take images in all weather conditions, has a repeat-cycle of 12 days, covers a large area and the data is available easily (Devanthry et al., 2016). Figure 1: An interferogram generated from two SAR images dated 04-01-2017 and 28-01-2017 (left) and its coherence image where the dark spots are the regions of low coherence and bright spots are the regions of high coherence (right) for the Erta Ale volcano. In this paper, the surface deformations are caused due to the injection of magma into the superficial parts of the earth’s crust. For this method, an interferogram generated through InSAR along with its associated coherence image is being used to compute a score value for images with deformation and non-deformation information.
In the areas where the land changes drastically, for example, due to vegetation growth the phase difference between neighboring pixels observed by InSAR appears random (incoherent), and therefore no meaningful information can be retrieved. To filter out these incoherent regions the associated coherence image of the interferogram are taken into account (Lu et al., 2010). A coherence is a measure of the quality of an interferogram and tells about the surface type such as vegetation or rock. It ranges from 0.0 to 1.0, where values closer to 0 states that the phase information in the interferogram contributes to noise and for values closer to 1 states that the phase information in the interferogram is correct. An interferogram and its associated coherence image are illustrated in Figure 1.
InSAR is a promising technology for monitoring the earth surface deformation related to some natural hazardous events, such as the earthquake, volcano eruption, land subsidence and landslide. (Devanthry et al., 2016) proposes the Persistent Scatterer Interferometry (PSI) for monitoring deformation using the Sentinel-1 data. For this approach, a large number of SAR images of the same region are used to measure the velocity of the deformation. This technique requires persistent scatters for detecting and measuring the deformations which are usually less for vegetated and steep areas. Hence, this method performs well for the detection of urban areas such as landslides but has a limitation for volcanoes due to the non-availability of large datasets and persistent scatters. (Casu et al., 2009) uses the DInSAR-Small BAseline Subset (SBAS) technique for retrieving surface deformation of volcanic areas. In this technique, it uses datasets consisting of multiple ascending and descending SAR images of the regions to generate times series and velocity maps for measuring the deformations. However, this method also requires large datasets to provide significant results and is expensive.
Despite several methods available for monitoring deformations none of the techniques so far are capable of detecting surface deformations if there is a limited dataset. Thus, this paper approaches to solve this problem by using a simple cost-effective technique for volcanoes where the gradient of the interferogram is calculated and weighted with the coherence image to remove low coherent regions from the interferogram and compute a score for the images where high score depicts the occurrence of deformation. The further sections emphasize the method and the results of this approach and finally, the conclusion and future works are discussed in Section 5. The method is divided into three submodules:
processing of phase image,
processing of coherence image
computation of score value.
The complete workflow is illustrated in Figure 2. First, the tif files of all the three volcanoes are loaded. These tif images are two channeled containing the phase and coherence information. After loading the tif images the phase (interferogram) and the coherence images are extracted from each tif file. The gradient for each phase image is calculated to acquire information about the change in surface.
A weight map is computed from the coherence image to remove the inconsistency in the phase image. For this, first, a Gaussian blur is applied to the coherence image. A Gaussian blur is a type of image blurring filter that uses a Gaussian function. With the help of Gaussian blur, the image noise (in this case vegetation, soil etc) is reduced to a great extent. Thereafter a simple thresholding is applied to the Gaussian blurred image to remove the incoherent regions completely and produce a weight image where the incoherent regions have a weight of zero. For this paper, a simple thresholding is used where values below 0.33 are mapped to 0 weight and rest all values act as actual weights.
A score value is computed for each weighted phase gradient (WPG) image. Score computation is as defined in Equation 5: Score = å(WPG) number of non-zero element of WPG (3) The score value is then normalized with a base image of each volcano. The image with no deformation information is selected as the base image. The score normalization is shown below in Equation 6: Normalized Score(Erta Ale/Kilauea/Piton) = Score(Erta Ale/Kilauea/Piton) – Score(Erta Ale/Kilauea/Piton base image)
Finally, the score values are plotted for each volcano type to generate a time series. Figure 3 illustrates the phase, the coherence, the weight map and the weighted phase gradient image for the case when no deformation has occurred and Figure 4 illustrates the phase, the coherence, the weight map and the weighted phase gradient image for the deformation.
As this paper, aims for a low-cost, easy to implement method it has used simple algorithms for applying thresholding and computing the score values. After, several tweaking of the blur parameter sigma and threshold (thresh) value the most optimal result is obtained with a sigma value equal to 12 and the threshold value equal to 0.33. This is compared in Figure 5 and 6 for Erta Ale. The time series for each of the volcano was generated to visualize the change in score value. Images with surface deformation information had a positive score whereas images with no deformation information had a negative score value.
In this paper, an easy to implement, low-cost approach is being proposed for detecting surface deformations due to the volcanoes. The score computation has been normalized which could automatically monitor new volcanoes on a global scale using the Sentinel data. Also, this algorithm is fast compared to PSI, SBAS which uses a large amount of data for processing. The use of satellite-based techniques saves a lot of time and cost compared to ground-based techniques. With the help of this method, various researchers and volcanologists can study the volcanic pattern of most tedious volcanoes as well and make better predictions of future activities and would thus save human and environmental losses. However, there are some shortcomings to this method as well. As seen in the results the algorithm fails to detect the atmospheric and topographic changes for the Erta Ale volcano. Also, there are false alarms for the Piton de la Fournaise volcano.
Thus, In the future, a better noise filtering technique could be adopted that could work for the complicated data like Piton de la Fournaise. Also, ways to detect atmospheric and topographic changes need to be researched. One approach could be to use a different weight computation like Min-Grad weighted algorithm or MFC weighted algorithm (Weike and Goulin, 2013).