Tracking Local Sea Ice Extent in the Beaufort Sea Using Distributed Acoustic Sensing and Machine Learning

Accurate determination of Arctic sea ice extent is critical for measuring climate change, understanding marine ecosystems, and navigating the Arctic basin for traditional native communities and industrial activity (Kwok and Untersteiner, 2011). Current constraints on Arctic sea ice mostly rely on satellite imagery, which is typically limited to daily updates and spatial resolution of a few to tens of kilometers (Fetterer et al., 2010; Meier et al., 2015, 2021). Estimating the extent of sea ice through satellite imagery can be challenging due to the complex composition of ice floes with varying thickness and sizes intermixed with water (Meier and Stewart, 2019). Another limitation is that different sea ice data products for the same region may be inconsistent, depending on resolution and their sensitivities to ice properties (Meier et al., 2015). Previous studies (Tsai and McNamara, 2011; Anthony et al., 2017; Baker et al., 2019) have shown that sea ice significantly attenuates the ocean microseism power observed with seismic stations in Antarctica and the Arctic, and these studies highlight the potential of microseism amplitudes to monitor sea ice. Therefore, arrays of seismic stations in polar regions could complement satellite imagery to better capture sea ice extent. However, the remoteness of the Arctic, difficult environmental conditions, and costs can make the deployment of numerous seismic stations unfeasible.

Distributed Acoustic Sensing (DAS) is an emerging technology that can effectively replace the necessity for large‐N seismic arrays in different environments. DAS uses a fiber‐optic cable to measure strain spatially resolved along the fiber, that is, repurposing the fiber strand as a dense seismic array (Lindsey and Martin, 2021). In this study, we interrogated a dark telecommunications fiber‐optic cable (a portion of a network owned by Quintillion Global) extending into the Beaufort Sea from Oliktok Point, Alaska (Fig. 1). Sea ice in the Beaufort Sea follows an annual cycle (Johnson and Eicken, 2016) starting with new ice forming near the shore around November, a stable ice period from February to June, a break‐up period around June or July, and an ice‐free period from July to October. The DAS data were recorded for a week during 9–15 July 2021 and another week during 10–16 November of the same year, specifically targeting time periods with transitional sea ice coverage. The DAS acquisiton parameters used for data collection were 10 m gauge length, 2 m spatial sampling, and 1000 Hz sampling rate. Figure S1, available in the supplemental material to this article, shows a snippet of DAS data. We use a Silixa iDAS interrogator that has a theoretical range of 40 km, and there is an optical repeater located at 42 km. We recorded data out to 37.4 km from the onshore interrogator spanning water depths of 0–20 m. Using unsupervised clustering of ambient noise, we show that DAS in polar seafloor environments captures sea ice extent with higher spatial and temporal resolution than existing satellite data products.

We use strain rate measurements spaced every 200 m (instead of the minimum available spacing of 2 m). We discard data from distances less than ∼0.5 km along the fiber, because this section is inside a conduit. The rest of the cable is directly buried. We also discard data from distances further than 35 km because of low signal‐to‐noise ratio due to distance from the interrogator. Figure 2 shows a summary of our workflow. We cut strain rate measurements along the cable into 30 min segments. We convert each 30 min segment into a scalogram using a Continuous Wavelet Transform (CWT) with a Morlet wavelet. The scalograms are limited to frequencies between 0.07 and 500 Hz. The scalograms are then saved as images. We generate 112,554 images for the two weeks of data. Different deep learning algorithms, such as Convolutional Neural Networks (CNNs), have been employed in audio classification tasks (Cummins et al., 2017; Amiriparian et al., 2019) or to identify marsquakes (Barkaoui et al., 2021). Seismograms resemble audio signals, because they are both continuous time‐series waveforms data albeit with different sampling rates (∼10 to 100 Hz vs. 44.1 kHz, respectively), meanwhile DAS have an intermediate sampling rate (kHz). The similarity between seismograms and audio signals, in addition to the success of Machine Learning (ML) with such datasets, lead us to test an unsupervised ML approach. Our goal is to reveal patterns in the DAS dataset without relying on previous information. In our case, we use InceptionV3 (Szegedy et al., 2016), a deep CNN, which was pretrained with the ImageNET database (Russakovsky et al., 2015), as a feature extractor for our scalograms.

A (MiniBatch) K‐means cluster (Sculley, 2010) analysis is performed using the features obtained from the CNN to identify groups of similar ambient seismic noise. We train our K‐means with a randomly chosen subset of 10,000 scalograms. At this stage, the number of clusters (K) and their physical meaning was unknown. We used the Calinski–Harabasz index (Caliński and Harabasz, 1974) to evaluate the optimal number of clusters. The Calinski–Harabasz index is a measure of how similar an object is to its own cluster (cohesion) compared to other clusters (separation). Such analysis indicates that there are two clusters in our dataset (K = 2). We obtain a binary classification of the ambient seismic noise for those two weeks of data with a temporal resolution of 30 min and a spatial resolution of 200 m. In the following section, we interpret the clusters and assign physical interpretations based on their spectral content.

In general, a cluster is a group of points that are similar to each other. In our case, a cluster is a group of scalograms with similar time–frequency information. We show, in Figure 3, the median Power Spectral Densities (PSDs) for each cluster. Cluster 1 is more energetic than cluster 2 for frequencies between ∼0.07 and 7 Hz. Cluster 1 displays two main energy peaks around 0.2 and 1 Hz, and a minor peak at 5 Hz, whereas cluster 2 does not show any similarly prominent peaks (see Fig. S2 for scalograms that belong to each cluster). Less prominent and closely spaced peaks between 7 and 10 Hz are common to both clusters’ median and first quartile lines (Fig. 3). We speculate that these common features are related to resonance in shallow seafloor sediments. Multiple studies (Sladen et al., 2019; Williams et al., 2019; Spica et al., 2020; Xiao et al., 2022) using ocean‐bottom DAS have observed different ocean–solid earth interactions, including ocean surface gravity waves or Scholte waves with peak frequencies around 0.05–0.2 Hz and 0.5–2 Hz, respectively. Cluster 1 displays the same PSD peaks, which indicates that their origin is related to wind‐driven ocean gravity waves. Similar PSD peaks have been observed (Stutzmann et al., 2009; Grob et al., 2011; Tsai and McNamara, 2011), with land seismometers near the shore, over multiple years during summer months in Antarctica and Alaska, when ocean microseism noise is strong, in contrast to the winter months when microseism noise is diminished by widespread sea ice.

Cluster 2’s median PSD displays no prominent peaks across most of the spectrum (0.07–10 Hz) and a gradual PSD increase from 10 to 500 Hz. We attribute the spectral behavior of cluster 2 to the presence of sea ice for two main reasons: (1) an attenuating material must be present to suppress the spectral peaks observed in cluster 1, that is, sea ice, or a material that prevents the excitation of oceanic waves, and (2) satellite imagery confirms some presence of sea ice during both the weeks (July and November). Sea ice attenuate waves for a large range of frequencies, including the primary microseism (0.05–0.1 Hz), secondary microseism (0.1–0.5 Hz), and other sources of noise such as lake microseisms (0.2–2 Hz) (e.g., Stutzmann et al., 2009; Grob et al., 2011; Tsai and McNamara, 2011; Anthony et al., 2017, 2018; Xu et al., 2017; Baker et al., 2019). The minor energy between 0.1 and 0.2 Hz might be related to ocean–land wave interactions that may not originate above cable and so cannot be fully attenuated by sea ice. The gradual PSD increase from 10 to 500 Hz observed in Figure 3 might be explained by noise levels increasing with distance or partially by instrumental effects (Lior et al., 2021). The higher amplitude of cluster 2’s median PSD in comparison with cluster 1 may suggest that the seafloor seismic signature of sea ice cover can include addition of high‐frequency noise (> 10 Hz) in addition to suppressing low‐frequency signals from ocean‐wave action. A more detailed analysis of these high‐frequency noise sources is out of the scope of this study. In the following section, we compare our cluster classification with the available satellite imagery.

We first describe DAS noise clustering as a proxy for spatiotemporal variations in sea ice extent and compare to more conventional satellite‐based constraints. Figure 4a,b shows the cluster classification in colours, that is, the inferred presence of sea ice (purple) and open water (orange). The horizontal axis represents the distance along the cable, and the vertical axis represents time in Figure 4a,b. Figure 4a shows, during July, the presence of a coastal area of open water surrounded by ice, that is, a polynya. Moreover, an abrupt change from cluster 1 to 2 (solid black line) for the week of July indicates a migrating sea ice edge. The sea ice edge (or the extent of the polynya) reaches the maximum along‐cable distance of 17.6 km on 11 July at 5:30 a.m. UTC and the minimum of 7.6 km, on 12 July at 1:30 a.m. UTC, implying that there could be changes in sea ice extent up to 10 km over hours or days. On average, the sea ice edge is located at 11.4 km (with a $σ=1.2km$⁠) along the cable for the entire week of observation in July. Figure 4b shows the cluster classification during the week from November. Figure 4b shows that early during 10 November the entire cable displays the behavior of cluster 1, that is, open water. In a time period of a few hours, we observe an almost complete change from cluster 1 to cluster 2 along the cable, indicating a change from open water to newly formed sea ice.

Complementary constraints on sea ice extent were collected from two satellite‐based sources: The Near‐Real‐Time Defense Meteorological Satellite Program (NRT DMSP) product (Meier et al., 2021) and the Multisensor Analyzed Sea Ice Extent (MASIE) product (Fetterer et al., 2010). We do not include other satellite data because it can be difficult to interpret, even for expert analysts, and might lead to ambiguous presence of sea ice (Meier et al., 2015). The NRT DMSP uses passive microwave imaging, whereas MASIE integrates multiple operational data streams including visible‐infrared, synthetic aperture radar, scatterometer, and passive microwave. MASIE provides estimates of daily sea ice extent using 1 km grid and a binary threshold of 40% sea ice concentration (white circles in bottom panels of Fig. 4). NRT DMSP provides estimated daily sea ice extent on a 25 km grid along with a continuous estimates of sea ice concentration in each grid cell (black squares with percentage numbers in bottom panels of Fig. 4). Data are available for both weeks from MASIE and for only one week (10–16 November) for NRT‐DMSP. Figure 4c–e shows the presence (or absence) of sea ice according to all the three datasets for three different times, highlighted by the arrows in Figure 4a,b. There is a close agreement between the sea ice estimates from MASIE and DAS noise classification during the week of July. Movie S1 shows the sea ice extent in time for the entire week in half‐hour intervals. Both datasets indicate the presence of sea ice between 7 and 15 km along the fiber‐optic cable. Some differences show up between sea ice edge observations for 13–15 July (see Fig. S3 for an example), for which DAS noise clustering indicates that the sea ice edge is at ∼12 km, whereas MASIE suggests that is located at ∼15 km. Such small differences may be expected, considering that MASIE’s sea ice concentration threshold is 40%; so if sea ice drops below that concentration in a 3 km segment, then it will be considered open water by MASIE.

The data from MASIE show that most of the path along the cable is covered by sea ice during the entire week of November (Fig. 4d,e). Meanwhile, the NRT DMSP indicates the presence of sea ice with low sea ice concentrations (< 10%) during 10 November and then sea ice concentration increasing to near 90% by 16 November, in agreement with our DAS observations. Movie S2 shows the entire progression of sea ice for the three datasets during November. DAS noise clustering suggests the presence of open water before mid‐day 10 November (Fig. 4d) with some ice in the first 10 km and then a change over a time period of a few hours to sea ice along most of the cable. We interpret the low sea ice concentration (from NRT DMSP) and our DAS observations as open water with discontinuous pieces of thin and newly formed sea ice that are insufficient to suppress wind‐driven ocean gravity waves. Later, the ocean above the cable begins to freeze outward from the shore. Movie S2 shows more clearly that according to our DAS observations the sea ice starts forming near the coast and then propagating outward, discarding the possibility of fragmented sea ice being pushed by wind toward the land.

There are some discrepancies between sea ice extent estimates from DAS and satellite‐derived products, which likely reflect differences in sensitivity to sea ice properties. Differences in satellite‐based sea ice products have been attributed to (Partington, 2000; Meier et al., 2015; Emery and Camps, 2017; Meier and Stewart, 2019): (1) different spatial resolution, (2) different sea ice concentration thresholds (15% vs. 40%), and/or (3) different sensitivity to melt, thin sea ice, or small floes depending on the instrument. Moreover, there may be outliers in DAS noise cluster classification (isolated spots in Fig. 4a,b) related to other transient noise sources such as marine mammals, icequakes, or anthropogenic activity (e.g., ships). The U.S. Geological Survey earthquake catalog reports no local earthquakes in our study region (Fig. 1); hence, we consider that local seismicity has a negligible influence on the noise classification results.

Ambient seismic noise (Grob et al., 2011; Tsai and McNamara, 2011; Anthony et al., 2017, 2018; Baker et al., 2019) has been studied by analyzing temporal changes in frequency content, and prior research found a clear anticorrelation between the microseism power and near‐coastal ice presence. In fact, most of those studies noticed that the microseism power is at its lowest point during winter months when the sea ice extent is the maximal. In addition, those studies were limited to a few seismic stations and provided no estimations of sea ice sensitivity of microseism due to the increasing distance from those seismic stations, which is what climate scientists are more interested in, that is, the sea ice coverage. An exception is the investigation by Cannata et al. (2019) for which they found that sea ice located as far as 1000 km modulates microseism amplitudes. Conversely, we observe that sea ice can change the seismic noise over distances as short as a few hundreds of meters. Such stark contrast is a consequence of the fiber‐optic cable that senses the poroelastic strains in the solid earth induced by the dynamic pressure field of ocean waves propagating above the cable, whereas seismometers only measure velocity or displacement at a given location. Other noise sources in the ocean include ships and marine animals that can generate signals in a broad frequency range, spanning from 0.01 to 100 kHz (Duarte et al., 2021), which overlaps with the maximum frequency we include in our analysis (500 Hz). The sea ice might add more localized events such as icequakes that have been observed with DAS elsewhere (Xie et al., 2023) and also emit energy at frequencies as high as hundreds of hertz. We plan to investigate these possible high‐frequency sources in the near future.

Unsupervised ML approaches have been used in numerous seismological applications (e.g., Seydoux et al., 2020; Jenkins et al., 2021; Steinmann et al., 2022; Hourcade et al., 2023). Our approach (Fig. 2) has the advantage of automatically extracting relevant features to identify the dominant noise types without any prior information. We tested our workflow with higher temporal resolution (5 min) and lower spatial resolution (500 m) with a subset of the original data and obtained similar results (Fig. S4). This suggests that the spatiotemporal resolution can be readily adapted to suit specific applications (e.g., local navigation vs. larger‐scale comparison with satellite imagery). We apply the Gradient‐weighted Class Activation Mapping (Grad‐CAM) algorithm (Selvaraju et al., 2017) to understand what features in the scalograms have greater importance for signal classification. Grad‐CAM highlighted low‐ and high‐frequency regions for scalograms that belong to cluster 1 and mostly high‐frequency regions (> 10 Hz) for scalograms that belong to cluster 2 (Fig. S5), suggesting that some relevant features in cluster 2 are present at high frequencies. Our study shows that DAS systems deployed in the Arctic are a valuable tool that can provide high spatiotemporal sea ice data products, but further testing is required to develop near‐real‐time implementations. Such near‐real‐time implementations would help to track sea ice more accurately, which may be especially important in a rapidly warming Arctic environment with evolving spatiotemporal patterns of sea ice coverage.

This observational study indicates that DAS can be used for automated tracking of local sea ice extent in the Arctic. We used an unsupervised ML approach to identify types of seismic noise in a polar marine seafloor environment (Fig. 2). We infer that the two dominant noises are related to the presence of sea ice and the presence of open water, which have distinct time–frequency patterns (Fig. 3). Figure 5 illustrates that the three main observations we inferred with DAS and satellite constraints are: (1) a polynya with an average position of 11.4 km along the fiber‐optic cable during July (Fig. 5a); both DAS and MASIE illustrate the sea ice edge around the same position. (2) An open‐water system with what we interpret to be thin pieces of floating ice during 10 November (before mid‐day); both NRT DMPS and DAS display signatures of open water, whereas MASIE suggests the presence of sea ice (Fig. 5b). We attribute such differences to thin discontinuous pieces of ice that could exceed MASIE’s sea ice concentration threshold without suppressing ocean waves. (3) The presence of sea ice as a result of a rapid refreeze event that occurred in a short time period, after which all three sea ice estimates agree (Fig. 5c). We recognize that there are some limitations to the application of DAS to track sea ice in Polar regions: (1) not the entire Arctic is covered with fiber‐optic cables (see Fig. S6), and there are no such cables in Antarctica. However, some key locations in the Arctic such as Hudson Bay, Labrador Sea, Greenland Sea, Barents Sea, Kara Sea, Laptev Sea, East Siberian Sea, Chukchi Sea, and Bering Sea, already have fiber‐optic cables deployed that, in theory, can be used to provide details about sea ice extent, but access to such cables is limited to telecommunications companies. (2) We infer the presence of sea ice along the cable, but we cannot determine how far the sea ice is present perpendicular to the cable. Such information may be difficult to obtain without the help of satellite constraints. (3) We do not know the exact minimum thresholds for the sea ice thickness and concentration to be detectable by our DAS workflow. Notwithstanding these limitations, the comparison between DAS noise clustering and sea ice constraints from satellite‐based sources suggests that DAS technology in Arctic environments has a great potential to constrain local sea ice extent by mapping where sea ice thickness and continuity are sufficient to diminish ocean waves.

We use sci‐kit learn (Pedregosa et al., 2011) for the clustering part (MiniBatchKmeans) and tensorflow (Abadi et al., 2015) for the convolutional neural network (CNN; InceptionV3). Figures 1 and 4 were made with PyGMT (Uieda et al., 2022). The supplemental material includes figures and movies that complement our observations. A part of the DAS data will be available at https://mhkdr.openei.org/submissions/438. A part of the code is available at https://github.com/andresp-wave/DAS_ice.git. Sea ice data used in this article came from the National Snow and Ice Data Center, and can be accessed in the two following links: https://nsidc.org/data/nsidc-0081/versions/2 and https://nsidc.org/data/g02186/versions/1. All websites were last accessed in July 2023.

The authors declare that there are no conflict of interests recorded.

The authors thank Editor‐in‐Chief Keith Koper, the associate editor Dominik Gräff, and an anonymous reviewer who provided valuable comments that improved the article. The authors thank Quintillion for fiber access and Silixa, LLC, for data collection and analysis advice. The authors would like to thank the University of New Mexico (UNM) Center for Advanced Research Computing, supported in part by the National Science Foundation (NSF; EAR‐2146272), for providing the high‐performance computing resources used in this article. This article was supported by the Laboratory Directed Research and Development program at Sandia National Laboratories (SNL)—a multimission laboratory managed and operated by National Technology and Engineering Solutions of Sandia LLC—a wholly owned subsidiary of Honeywell International Inc. for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE‐NA0003525. This article describes objective technical results and analysis.

Any subjective views or opinions that might be expressed in the article do not necessarily represent the views of the U.S. Department of Energy or the U.S. Government.