Cross-domain fusion in smart seafloor sensor networks

The 2018 volcanic sector collapse of Anak Krakatau (Indonesia) and the 2022 explosive eruption of the volcano Hunga Tonga-Hunga Ha’apai (Tonga) triggered tsunamis that caused severe damage to nearby coastal settlements [2, 3]. Most coastal regions surrounding the Pacific and Indian Oceans are equipped with early-warning systems for tsunamis triggered by major earthquakes. However, neither submarine volcanic eruptions nor the associated tsunamis are part of monitoring and early warning systems, leaving coastal societies unprepared. Unlike earthquakes, volcanic eruptions and landslides are generally preceded by a wide range of precursory signs ranging from deformation, heat and gas emissions or seismicity, which in the case of onshore volcanoes can be monitored on-site or by remote sensing. Satellite-based remote sensing has become one of the most important tools in monitoring (especially remote) volcanoes [4] and not only allows monitoring the course of an eruption, but also precursory processes like gas and ash emissions or deformation using Interferometric Synthetic Aperture Radar (InSar) data seen for the 2018 Anak Kraktau eruption [5, 6] or the 2021 Cubre Vieja (La Palma, Canary Islands) eruptions [7]. However, some crucial volcanological measurements like seismicity or the analysis of gas compositions can only be done with ground-based sensors, while the submerged part of marine volcanoes cannot be monitored by satellite at all. To achieve a shoreline-crossing monitoring of marine volcanoes, it is necessary to combine both area-wide measurements using remote sensing techniques with point measurements using stationary seafloor sensors (lander). Monitoring with remote sensing often has large temporal gaps in the coverage of a specific target, while point measurements often lack sufficient coverage to unambiguously detect processes. However, the shortcomings of each measurement approach can be effectively compensated by the other, but requires the cross-domain fusion of complex data streams (Fig. 1). Concepts to transfer data between the marine and the subaerial parts of integrated networks can be realized with buoys or autonomous underwater vehicles acting as relays [8]. The main aim of volcano monitoring is to detect potential hazards, and thus information from the submarine domain needs to be transferred onshore. This requires communications strategies that reduce the amount of transferred data either in predefined intervals or when changes in the system occur, which requires robust pattern recognition on-device or within the underwater network.

Carbon Capture and Storage (CCS) is one of the key technologies for mitigating future CO₂ emissions, and its industrial-scale implementation is part of every mission reduction path to limit global warming to 2 or even 1.5 °C defined by the International Panel of Climate Change in their 2018 special report [9]. The most promising geological storage in Europe is represented by exploited hydrocarbon reservoirs and marine aquifers. Marine aquifers in the Norwegian sector of the North Sea alone have a storage capacity of 70 gigatons [10], and these are the storage formation of the first and at the same time only actively operating submarine CCS as site Sleipner, where CO₂ is injected in the sandstones of the Utsira Formation ~ 900 m below seafloor [11]. Peak CO₂ injection rates were 1 Mt per year, which is only a fraction of the up to 10 Gt per year required in the IPCC to achieve the 1.5 °C path [9]. Extending injection rates towards a degree sufficient to contribute to global CO₂ emission budget requires the implementation of novel, cost-efficient monitoring strategies since current reservoir performance and leakage detection monitoring builds on three-dimensional (3D) seismic surveying, which can only be repeated in multi-year intervals and are difficult to scale. However, cost-efficient fiber-optic cable and seafloor lander-based monitoring techniques are currently in development [12]. These allow continuous monitoring of seafloor deformation as proxy for fluid migration in the underground, but require repeated ground truthing using seismic data and need to be calibrated by geomechanical and fluid flow simulations [12]. These data domains are highly uncorrelated and require cross-domain fusion to enable actual data fusion and mining. Further, the limited communication links of the devices practically enforce on-device analytics to reduce the need for communication. Monitoring strategies for CO₂ storage sites needs to be able to detect long-term changes (e.g., increase of tilt or uplift over time) as well as sudden events indicating potential problems in the integrity of the storage formation (e.g., shallow seismicity indicating hydrofracturing or changes in the water chemistry). In the following, we discuss how intelligent Underwater Sensor Networks can address such challenges.

Wireless broadband communication is an important cornerstone of modern network computing and is not possible underwater since high-frequency electromagnetic waves are absorbed by seawater due to its high electrical conductivity. Therefore, underwater communication is generally limited to acoustic communication, which operates at frequencies of between 10 and 400 kHz with data transfer rates of 10s of kbit/s over some hundred meters. Present-day 5G mobile communication has data transfer rates, which are about a million times faster using frequencies between 24 and 54 GHz, enabling 10s of Gbit/s over several kilometers. Acoustic waves (speed of sound of water: ~ 1,500 m/s) travel with velocities a million times slower than electromagnetic waves (speed of light: ~3 × 108 m/s), adding a significant latency to acoustic communication [13]. Adding to these limitations, acoustic broadcasting has high energy consumption and thus requires high battery capacities, which limit the size and cost reduction potential of underwater network nodes for IoUT applications [13].

To benefit from the advances of reducing the required dimensions for seafloor devices and at the same time establishing communication within underwater networks, the required amount of transferred data within wireless underwater networks needs to be minimized. Underwater sensors record large amounts of data, while having only limited communication capabilities. A hydrophone, for example, samples at 4 kHz with 24 bits, resulting in a data rate of 12 kBit/s. While this data rate can easily be handled by today’s cellular technologies such as 4G and 5G, underwater communication is limited to a few bits per second due to the above-mentioned challenge. Moreover, underwater communication is highly energy intense [13].

Sensor nodes on the seafloor collect multivariate time-series data: Each node can be equipped with a vast sensor array including hydrophone, accelerometer, gyroscope, pressure, temperature sensors. Often operating as continuous, long-term deployments, sensors provide data at high resolution in time and for a long duration. However, due to its limited scope, a sensor only covers a limited geographic space. Remote sensing, in contrast, covers a vast geographic area. Its resolution in time is, however, limited to the times the satellite or airplane is in the vicinity of the area of interest. Remote sensing data are commonly in the form of a 2D or 3D image. Due to their different modalities, time series and image data cannot be directly fused. Instead, each need to be analyzed to identify patterns and events of interest that can serve as a common base. The resulting patterns can then be fused and form the basis of further data analytics.

The overall application setting is a distributed setting with a built-in hierarchy. The underwater sensor nodes form a distributed system with strong resource constraints in terms of compute-power, communication capabilities and energy. In contrast, the remote sensing data is commonly available at a place with vast compute and communication capabilities. Due to limited communication links, data needs to be aggregated with the network into compact higher-level representations, i.e., patterns [14]. A pattern is a higher-level representation of an event of interest, such as an underwater sensor detecting an earthquake [15]. Patterns can be identified from data of one or more sensors in close proximity. Moreover, communicating compact patterns instead of raw data limits the load on the constrained communication system (see Fig. 1).

Event or pattern recognition may also be applied to data from other sensor types, e.g., cameras [16]. When working with distributed, resource-constraint networks, communication need to be reduced to a bare minimum. This can be done by simply communicating changes in the environment-based predefined parameters. All mentioned techniques can benefit from recent approaches in machine learning, which allow pre-defining neural networks to be implemented on the network nodes itself. Fusion of raw or processed data or ideally of already classified events may be implemented decentralized within the network or on a dedicated network node depending on the requirements of the specific network. To address these challenges, new methodologies to individually analyze each data stream to identify patterns and events of interest need to be devised. Building on neural networks and distributed computing, these methodologies will need to be resource-efficient so that they can be deployed on battery-driven sensor nodes that will be deployed over multiple years on the ocean floor. The patterns and events of interest then serve as a common base and can be fused and form the basis of further data analytics.