PaperlayerView exact revision on arXiv

Source-linked AI summary

Internet of Underwater Things and Big Marine Data Analytics -- A Comprehensive Survey

Mohammad Jahanbakht, Wei Xiang, Lajos Hanzo, Mostafa Rahimi Azghadi

arXiv:2012.06712v2cs.NI

TL;DR

IoUT and Big Marine Data remain challenging because underwater applications are underexplored and generate difficult-to-process data. This survey synthesizes communications, data processing, and machine-learning approaches, highlighting open research issues and future directions.

Abstract

from arXiv · show

The Internet of Underwater Things (IoUT) is an emerging communication ecosystem developed for connecting underwater objects in maritime and underwater environments. The IoUT technology is intricately linked with intelligent boats and ships, smart shores and oceans, automatic marine transportations, positioning and navigation, underwater exploration, disaster prediction and prevention, as well as with intelligent monitoring and security. The IoUT has an influence at various scales ranging from a small scientific observatory, to a midsized harbor, and to covering global oceanic trade. The network architecture of IoUT is intrinsically heterogeneous and should be sufficiently resilient to operate in harsh environments. This creates major challenges in terms of underwater communications, whilst relying on limited energy resources. Additionally, the volume, velocity, and variety of data produced by sensors, hydrophones, and cameras in IoUT is enormous, giving rise to the concept of Big Marine Data (BMD), which has its own processing challenges. Hence, conventional data processing techniques will falter, and bespoke Machine Learning (ML) solutions have to be employed for automatically learning the specific BMD behavior and features facilitating knowledge extraction and decision support. The motivation of this paper is to comprehensively survey the IoUT, BMD, and their synthesis. It also aims for exploring the nexus of BMD with ML. We set out from underwater data collection and then discuss the family of IoUT data communication techniques with an emphasis on the state-of-the-art research challenges. We then review the suite of ML solutions suitable for BMD handling and analytics. We treat the subject deductively from an educational perspective, critically appraising the material surveyed.

NOMENCLATURE … B. IoUT Communications

The paper surveys the Internet of Underwater Things (IoUT), Big Marine Data (BMD), and machine-learning analytics, progressing from underwater data collection and communications to processing challenges and future directions. It addresses IoUT’s heterogeneous, energy-constrained environment and the need for specialized analytics for massive, diverse marine data.

NOMENCLATURE

  • The nomenclature defines the abbreviations used throughout the IoUT, BMD, communications, and machine-learning discussion.Key terms include IoUT, BMD, UWSN, ML, DL, CNN, RNN, SVM, and QoS.Source: NOMENCLATURE · p. 1

I. INTRODUCTION

  • IoUT research remains comparatively unexplored despite the importance of oceans, which cover over 70% of Earth’s surface and support 90% of international trade.The paper identifies underwater technologies as important for understanding, controlling, and interacting with marine environments.Source: §I · p. 2
  • IoUT differs from terrestrial IoT because underwater communication environments, computational limitations, and constrained energy resources create distinct technical challenges.The survey follows a progression from IoUT data collection and networking to BMD analytics and processing, covering sensors, communications, machine learning, and future challenges.Source: §I · pp. 2–4
  • The paper’s distinctive contribution is a comprehensive survey combining IoUT, BMD analytics, and recent machine/deep-learning approaches.It addresses a gap left by prior surveys that treated IoUT, BMD, or terrestrial IoT analytics separately,,,, –, –.Source: §I · p. 3

II. INTERNET OF UNDERWATER THINGS

  • IoUT is defined as globally interconnected networks of digitally identified underwater objects that follow reference-model protocols such as TCP/IP or OSI.The survey examines IoUT objects, underwater communications, and network standards and protocols.Source: §II · p. 4

A. Underwater Sensors and Devices for Data Collection

  • Sparse deployment and limited coverage make low-energy sensors essential for monitoring vast, hostile, and dynamic underwater environments.Existing small-scale UWSNs and research vessels cannot cover seas spanning thousands of square kilometers, whereas Internet-connected sensor networks can provide historical and real-time measurements for prediction and policy decisions,.Source: A. Underwater Sensors and Devices for… · pp. 4–5
  • IoUT data collection combines endpoint devices, mid-layer redistribution nodes, and overwater or land-side sink nodes, with sensors generating continuous BMD for transfer, storage, and processing.Endpoint devices include sensors, cameras, hydrophones, storage chips, actuators, and tags; mid-layer nodes include modems, gateways, repeaters, and relays; sinks include buoys, ships, satellites, and shore stations,.Source: A. Underwater Sensors and Devices for… · p. 5

B. IoUT Communications

  • Underwater communications are constrained by acoustic propagation delay, fading, narrow bandwidth, high attenuation, and low channel capacity, motivating combinations of acoustic, electromagnetic, and optical technologies.Signal attenuation directly constrains maximum reliable data rate and communication distance, making reliable long-range underwater transport a major IoUT challenge.Source: B. IoUT Communications · p. 5, A. Underwater Sensors and Devices for… · p. 5

1) Electromagnetic signal attenuation: … C. Improving IoUT Link Reliability

The survey explains how seawater attenuation and environmental variability constrain IoUT communication, making link reliability dependent on channel-specific metrics and technology tradeoffs. It concludes that long-range links favor low-rate acoustics, whereas short-range links can use high-rate optics, motivating routing and hop-count improvements.

1) Electromagnetic signal attenuation:

  • Electromagnetic waves attenuate strongly in conductive seawater, so only low-frequency signals travel long distances, with reduced channel capacity and larger antennas as frequency decreases.Penetration depth decreases with increasing conductivity and frequency.Source: 1) Electromagnetic signal attenuation · p. 5

2) Acoustic signal attenuation:

  • Acoustic links dominate IoUT communications because electromagnetic propagation is hostile underwater, although acoustic channels also have narrow bandwidth and lower low-frequency attenuation.Acoustic attenuation depends on frequency, pH, temperature, salinity, and depth.Source: 2) Acoustic signal attenuation · p. 5

3) Optical signal attenuation:

  • Underwater optical attenuation arises mainly from absorption and scattering, with radiative-transfer solvers offering precision but high complexity and Monte Carlo models offering speed but lower accuracy,,.Finite-difference and improved-integral methods increase accuracy but remain computationally expensive.Source: 3) Optical signal attenuation · pp. 5, 7

C. Improving IoUT Link Reliability

  • Dynamic tides, pressure, temperature, sediments, and chemical changes distort amplitude, disperse frequency, refract speed, and create delay jitter, rendering underwater channels unreliable,.The survey therefore calls for reliability metrics to quantify and optimize IoUT quality of service.Source: C. Improving IoUT Link Reliability · p. 7
  • Link reliability can be measured through hardware SNR-based metrics or software delivery metrics, including BRR/BER, PRR, RNP, and ETX –.RNP is measured at the transmitter, whereas BER is measured at the receiver; ETX is the reciprocal of PRR.Source: C. Improving IoUT Link Reliability · p. 7
  • BER, RNP, and ETX depend on data rate, transmitted power, and transceiver distance, so reliable BMD delivery requires sufficient received power despite random background noise.The survey identifies ad hoc routing improvements and hop-count optimization as techniques for boosting software- and hardware-based link reliability.Source: C. Improving IoUT Link Reliability · pp. 7–8
  • Reliable underwater wireless communication is restricted to low-data-rate acoustic waves for long distances or high-data-rate optical rays for short distances, while short-distance electromagnetic waves are substantially outperformed [45].The comparison attributes these constraints to attenuation, bandwidth, and transmission-range differences across technologies.Source: C. Improving IoUT Link Reliability · pp. 7–8, Table III · p. 8

1) Routing improvement: · 2) Hop-count optimization: · D. IoUT Network Architecture

The section surveys reliable IoUT routing through clustered forwarding, relay selection, and hop-count optimization, then maps IoUT components and protocol requirements onto a layered TCP/IP architecture.

1) Routing improvement:

  • Clustered routing combines location-free, energy-based node grouping with store-and-forward buffering to improve SNR reliability, reduce retransmissions, balance load, and transmit fewer packets.The topology uses relay nodes and cluster-heads to forward data between sub-regions and upper cluster coordinators.Source: 1) Routing improvement · pp. 8–9, Figure 3 · p. 8
  • Tran-Dang et al. independently select relays and cluster-heads using SNR, hop-count, and time-of-arrival estimates that are repeatedly updated for changing underwater conditions.The reported simulations still require experimental fieldwork for validation.Source: 1) Routing improvement · p. 9

2) Hop-count optimization:

  • Hop-count optimization trades fewer relays against longer transmission distances, because greater range reduces bandwidth and raises power consumption under minimum-SNR delivery.Hop-count denotes the number of intermediate devices conveying data between successive sources and destinations.Source: 2) Hop-count optimization · p. 9
  • Li et al. use relay nodes along underwater links to reduce hop distance, offering an energy-efficient approach for sustainable high-data-rate transmission under power-bandwidth-range constraints.The scheme addresses battery-constrained nodes and their SNR reliability requirements.Source: 2) Hop-count optimization · p. 9

D. IoUT Network Architecture

  • IoUT architecture can be organized through the ubiquitous five-layer TCP/IP model: application, transport, network, data link, and physical layers.A four-layer variant combines the data-link and physical layers into a network-access layer.Source: D. IoUT Network Architecture · p. 9
  • Table IV maps land-side computers and underwater sensors to application endpoints while representing sink nodes, physical channels, underwater mid-layer nodes, and underwater endpoints across the IoUT architecture [61].The table uses layer names rather than standard numeric layer identifiers.Source: D. IoUT Network Architecture · p. 9, Table IV · p. 9
  • IoUT protocol stacks may combine multiple coexisting protocols to connect IoT and Web systems while addressing latency, coherence-time packet sizing, attenuation, application-specific bit errors, low bandwidth, and security threats.Protocols specify data formats, syntax, synchronization, security, and error control across layers.Source: D. IoUT Network Architecture · p. 10

1) Underwater application layer: … 5) Physical layer in IoUT:

The IoUT protocol stack assigns distinct responsibilities across application, transport, network, data-link, and physical layers, but must accommodate limited bandwidth, energy, intermittent connectivity, and harsh propagation conditions. Consequently, underwater systems favor lightweight protocols, energy-aware coordination, specialized routing and MAC schemes, multimodal communications, and analytical channel modeling.

1) Underwater application layer:

  • Application-layer nodes identify endpoints, collect and process sensor data, deliver commands, and may operate controllable actuators, while suitable protocols avoid unnecessary overhead.Examples include IMC, SLIP, XML, MQTT, DDS, Modbus, and Telnet; Telnet’s simple 8-bit, text-oriented structure suits IoUT applications.Source: 1) Underwater application layer · p. 10

2) Underwater transport layer:

  • Transport-layer protocols packetize endpoint data and manage ordering and loss; UDP is preferred over TCP when some packet loss is acceptable because it avoids acknowledgments and overhead.DCCP additionally provides congestion notification and control for UWSNs, while multimodal nodes can support acoustic, electromagnetic, and optical technologies.Source: 2) Underwater transport layer · p. 10

3) Underwater network layer:

  • The network layer handles bidirectional packet forwarding, protocol translation, Internet connectivity, and routing, with IPv4 favored over verbose IPv6 and ECN recommended when DCCP is used.Mid-layer nodes extend communication range under long-distance or high-attenuation conditions, while energy management remains necessary because underwater nodes are difficult to recharge.Source: 3) Underwater network layer · pp. 10–11
  • Cross-layer allocation of routing and scheduling resources achieved 66% less energy consumption per successfully transmitted bit, but the single-hop design leaves multihop delay–power trade-offs unresolved.The strategy prioritizes nodes and sinks with longer paths or less efficient physical layers, requiring further investigation of multihop operation.Source: 3) Underwater network layer · p. 11

4) Underwater data link layer:

  • The data-link layer converts frames and electronic signals and coordinates shared-channel access; collision-free MAC protocols generally outperform contention-based schemes in underwater acoustic networks.Underwater MAC designs must minimize frame size under limited bandwidth and energy, while specialized approaches address energy-aware scheduling, intermittent connectivity, and video transmission.Source: 4) Underwater data link layer · p. 11

5) Physical layer in IoUT:

  • Wired physical links support energy transfer and narrow-bandwidth sensing or wideband observatory data but suffer installation, maintenance, mobility, security, and environmental drawbacks.Fiber-optic systems support wideband image and video transmission, while underwater optical-fiber infrastructure also carries intercontinental Internet traffic.Source: 5) Physical layer in IoUT · pp. 11–12
  • Wireless underwater propagation is impaired by environmental conditions, with shallow water dominated by surface noise and multipath fading and deep water by attenuation, narrow bandwidth, and temperature sensitivity.Overwater electromagnetic protocols have very short underwater ranges, whereas acoustic waves are the only viable option for long-distance underwater wireless communication.Source: 5) Physical layer in IoUT · p. 12
  • Reliable physical-layer optimization requires analytical or numerical transmission evaluation using an appropriate underwater channel model to guide IoUT communication-system design.The channel model is presented as the basis for the simulation and optimization discussed in the following section.Source: 5) Physical layer in IoUT · p. 12

E. Underwater Channel Modeling · F. IoUT Network Topologies · G. Underwater Network Simulation Tools

The paper presents use-case-specific underwater channel models, topology choices, and simulation/emulation tools as foundations for designing, optimizing, and validating IoUT networks before deployment. It emphasizes that model simplifications and incomplete environmental inputs limit accuracy, requiring experimental validation.

E. Underwater Channel Modeling

  • Use-case-specific channel models predict transmission performance, optimize node placement, and reduce energy consumption before IoUT deployment.No universally accurate model applies across communication technologies; simplifications must match the application, such as neglecting salinity for acoustic but not electromagnetic propagation,.Source: E. Underwater Channel Modeling · p. 13
  • Dedicated acoustic models combine features such as layered water, scattering particles, and seabed slopes, while trading environmental fidelity against complexity and accuracy –.The common formulation superposes propagation rays, with statistical coefficients representing scattering, Doppler shifts, node uncertainty, received-power changes, and dynamic seabed topology.Source: E. Underwater Channel Modeling · p. 13
  • Channel modeling forms the backbone of IoUT simulation, linking physical-layer behavior to network-topology design and parameter optimization –.Line-of-sight optical channels are addressed separately, whereas optical-fiber modeling is outside the article’s scope,.Source: E. Underwater Channel Modeling · pp. 13–14

F. IoUT Network Topologies

  • Because underwater bandwidth is limited and energy harvesting is difficult, conventional tree and mesh topologies dominate: trees suit small one-way networks, while meshes support sophisticated high-traffic networks,,,,,,.Route reliability is quantified by overall bit error rate, enabling topology selection against an application’s target,,,,, [35],,.Source: F. IoUT Network Topologies · p. 14
  • Topology, multilayer-network, and protocol design are difficult to test and deploy under harsh underwater conditions, motivating simulation tools for iterative design and layer-wise testing.TCP/IP-based gateway-enabled applications can use known topologies, with tree and mesh identified as especially promising,.Source: F. IoUT Network Topologies · p. 14

G. Underwater Network Simulation Tools

  • Discrete-event simulators model underwater channels and protocols through repeatable event-triggered systems, supporting topology arrangement, traffic and routing, heterogeneous technologies, protocol testing, and performance analysis,.The ns family and extensions including AQUASIM, SUNSET, WOSS, and DESERT Underwater provide underwater channel modeling, acoustic transmission, and relevant protocols.Source: G. Underwater Network Simulation Tools · p. 14
  • Emulation after simulation is recommended before deployment because it mirrors real-network behavior, enables close-to-real-life troubleshooting, and improves confidence in IoUT quality.Except AQUASIM and OMNeT++, the listed tools support IoUT emulation with hardware and software solutions.Source: G. Underwater Network Simulation Tools · p. 14
  • Simulation accuracy depends on input detail and environmental realism; therefore, modeling results should be validated through experimental fieldwork, while security-specific tools assess remote IoUT networks’ cyberattack resistance.Uncertain factors include noise, distortion, electronics, seabed topography, currents, microorganisms, and species movement; NeSSi2 supports security-threat evaluation.Source: G. Underwater Network Simulation Tools · pp. 14, 16

H. IoUT Network Security … I. Software-Defined Networks in IoUT

IoUT security must address eavesdropping, unauthorized access, and active data manipulation despite mobile nodes, constrained resources, and limited energy. Cryptographic, covert-communication, secure-protocol, and SDN approaches improve protection, adaptability, and centralized management, although some hybrid designs require practical validation.

H. IoUT Network Security

  • IoUT networks face eavesdropping and unauthorized access because unattended nodes and widely broadcast physical channels expose sensitive infrastructures such as harbor security and coastal defense.Threats include passive data extraction and active injection, alteration, or deletion through malicious nodes.Source: H. IoUT Network Security · p. 16
  • Authenticated data access can provide confidentiality, integrity, freshness, user authentication, and non-repudiation, but mobility, communication and computation constraints, and energy shortages complicate implementation.Cryptographic primitives and secure communications are proposed to address these constraints.Source: H. IoUT Network Security · p. 16

1) Cryptography:

  • Symmetric encryption is preferred where possible in resource-constrained underwater environments because shorter keys avoid public-key broadcasting and sophisticated algorithm execution.Suitable algorithms include Blowfish, Twofish, Threefish, AES, RC5, and RC6; ECC is recommended for asymmetric cryptography because it is energy-efficient and uses smaller equivalent-security keys.Source: 1) Cryptography · p. 16
  • Physical-layer key generation avoids channel broadcasting and addresses IoUT key-sharing constraints, with a multiparty underwater acoustic scheme designed for multipath and Doppler effects.Simulations in demonstrate secure communication against key leakage during active attacks, whereas classical protocols impose infrastructure, complexity, and resource demands.Source: 1) Cryptography · p. 16

2) Secure communications:

  • Covert communications reduce eavesdropping by lowering average propagation-medium SNR through beamforming, frequency hopping, spread spectrum, code division multiplexing, and analog network coding.Camouflaged transmissions can shape signals to resemble killer-whale voices or ship-radiated noise, with studies verifying feasibility in IoUT.Source: 2) Secure communications · p. 16
  • Secure protocols span application, transport, network, data-link, and physical layers, including RADIUS, Diameter, SSL, TLS, IPSec, SeFLOOD, RPR, R-CARP, EAP, EAPOL, and MACsec.These distributed controls motivate centralized SDN management for stronger IoUT security.Source: 2) Secure communications · p. 17

I. Software-Defined Networks in IoUT

  • In IoUT, SDNs replace locally configured, node-level routing decisions with centralized control nodes that continuously monitor traffic flow and manage network configuration to improve overall performance.They are programmable, hardware-independent, automatically manageable, reconfigurable, multi-modal, scalable, and adaptable in size and topology.Source: §I · p. 17
  • SDNs improve IoUT security through global monitoring, enabling centralized controllers to detect irregular behavior and suspicious transmissions and isolate or remediate untrusted nodes.Their global system view provides better protection against malicious attacks than localized management.Source: §I · p. 17

J. Edge Computing in IoUT … 2) Data aggregation:

The paper surveys IoUT architecture and edge computing, then defines Big Marine Data and its acquisition pipeline, emphasizing communication constraints, machine-learning data sources, and aggregation trade-offs. It argues that distributed processing and specialized acquisition and aggregation methods are central to handling underwater data efficiently.

J. Edge Computing in IoUT

  • Edge computing moves IoUT processing toward underwater endpoints, mid-layer nodes, and acquisition tools, reducing data transfer, latency, and centralized decision dependence –.This is especially valuable because underwater communication is hostile and conventional cloud or local-server processing requires repeated data shuttling.Source: J. Edge Computing in IoUT · p. 18
  • IoUT edge computing remains constrained by limited device energy and the difficulty of delivering sustainable power to underwater processing units.Mobile Cloud Computing offers scarce-resource applications rich cloud resources but is poorly suited to IoUT because of long propagation distances, narrow bandwidth, limited energy, and Internet-only access,.Source: J. Edge Computing in IoUT · p. 18

K. Section Summary

  • The section concludes that edge computing may be more important for IoUT than conventional IoT because underwater communications are more challenging.The broader IoUT discussion also covers sensors, communication reliability, architectures, network modeling, topology, simulation, security, and software-defined networking.Source: K. Section Summary · p. 19

III. BIG MARINE DATA

  • Big Marine Data comprises vast, heterogeneous marine datasets spanning long temporal and broad spatial coverage, diverse sources, and multidisciplinary physical, chemical, biological, environmental, and economic types.Its system stages are acquisition; transportation and security; storage and privacy; special-purpose processing; and exploitation and leveraging [32].Source: §III · p. 19

A. Marine Data Acquisition

  • Marine data acquisition is organized into data gathering, data aggregation, and data fusion, using vehicles equipped with primary sources such as cameras, hydrophones, and sensors.The paper presents acquisition tools, localization and tracking methods, and ready-to-use ocean sensory, imagery, video, and geographic databases.Source: A. Marine Data Acquisition · p. 19, 1) Data gathering · p. 19, §III · p. 19

1) Data gathering:

  • About 95% of relevant ML publications used unmanned ROVs and AUVs together, while more than 50% of underwater ML research occurred after 2014.The post-2014 increase excluded human-occupied vehicles, plausibly reflecting wider adoption of automated and remote methods and reduced use of costly manned vehicles.Source: 1) Data gathering · p. 19, Table VII · p. 20

2) Data aggregation:

  • Data aggregation summarizes raw measurements through methods such as down-sampling and linear regression before higher-level computation or transmission over band-limited channels.In UWSNs it can reduce bandwidth use and network-level energy, but may increase node processing energy, latency, and accuracy loss.Source: 2) Data aggregation · pp. 19–20
  • LEACH, PEGASIS, HEED, and APTEEN can be adapted for UWSNs, while edge computing can mitigate aggregation’s heavy computation and energy demands –.These approaches address the trade-off between low processing power and high communication energy without eliminating aggregation’s latency and accuracy concerns.Source: 2) Data aggregation · p. 20

3) Data fusion: … 2) Two-dimensional underwater image data:

The section surveys IoUT data fusion and marine data modalities, covering heterogeneous-data integration, sensor time series, underwater imagery, and the physical constraints shaping their collection and processing.

3) Data fusion:

  • IoUT data fusion integrates heterogeneous sensor, audio, video, and command data using edge devices and cloud servers to obtain more consistent information than individual databases.Design choices concern fusion location, abstraction level, and overlap among original datasets.Source: 3) Data fusion · p. 20
  • Fusion may be centralized or distributed, with low-level fusion generally unsuitable for resource-constrained BMD except at the edge, while medium- and high-level fusion support sophisticated models,.Cooperative fusion of different data types is preferred over redundant or complementary overlap because underwater transmission is expensive.Source: 3) Data fusion · pp. 20–21

B. Marine Data Classification

  • IoUT marine data are primarily acquired by cameras and sensors integrated into wired or wireless sensor networks, whose datasets are classified by dimensionality.These networks form a predominant component of the overall IoUT infrastructure.Source: B. Marine Data Classification · p. 21

1) One-dimensional time-series marine data:

  • One-dimensional sensor and nano-sensor time series support long-term, large-scale environmental monitoring, with nano-sensors favored for low power consumption,.Applications include seawater-quality monitoring, environmental sensing, ecosystem observation, and anomaly detection using ML, including tropical-cyclone prediction,.Source: 1) One-dimensional time-series marine… · p. 21
  • Observatory stations provide globally accessible BMD through active, continuously updated projects, while legacy datasets such as POSEIDON and JGOFS remain downloadable despite discontinued support.Observatory measurements can contain environmental noise, outliers, misread values, and missing quantities requiring correction techniques.Source: 1) One-dimensional time-series marine… · p. 21

2) Two-dimensional underwater image data:

  • Two-dimensional underwater image databases include live observatory imagery and labeled still-image datasets suitable for supervised ML, although only a few receive continuous updates.The imagery supports ecological counting and tracking as well as ML training and performance verification.Source: 2) Two-dimensional underwater image data · pp. 21–22
  • Active imaging improves image quality with controlled artificial light but incurs high energy loss, reduced portability, and uneven intensity and color; backscatter and forward scatter are major optical artifacts.Forward and back scattering blur images increasingly with camera-to-object distance, motivating deblurring methods [49].Source: 2) Two-dimensional underwater image data · pp. 22–24
  • Passive imaging consumes less power but suffers limited visibility, low contrast, and color distortion, making it inefficient below roughly 10 meters and necessitating artificial light for deep-water exploration.Red light attenuates below 10 m, while blue wavelengths penetrate farther but produce low-contrast bluish images.Source: 2) Two-dimensional underwater image data · p. 24

3) Three-dimensional underwater video data: · C. Distributed and Cloud-based BMD Processing

Underwater video processing is constrained by extreme data volumes, manual annotation costs, communication limits, and depth-imaging artifacts, motivating automatic analysis and cloud or distributed platforms. These platforms support IoUT big marine data processing through Apache frameworks and cloud service models, each with distinct capabilities and limitations.

3) Three-dimensional underwater video data:

  • A camera recording at 1 frame/sec can generate 30 million frames, or almost 3 TB, annually, with total volume multiplying across cameras.Underwater video streaming also faces constrained communications and imaging difficulties, including contrast, color distortion, visibility, backscatter, and forward scatter.Source: 3) Three-dimensional underwater video… · p. 24
  • 15 minutes of expert annotation are required per video minute, implying approximately 10,800 man-hours to analyze one camera’s monthly output and motivating automatic processing.Automatic methods support applications including visible-light and sonar tracking, underwater photo mosaicing, and marine-life studies,,,.Source: 3) Three-dimensional underwater video… · p. 24
  • Depth-based video combines optical cameras, acoustic arrays, ToF sensors, or lasers to enable 3D reconstruction, tracking, sizing, counting, and improved 2D segmentation and classification -.It is deployed on underwater vehicles, tethered platforms, towed systems, and baited stations, -.Source: 3) Three-dimensional underwater video… · p. 24
  • Acoustic depth imaging suffers variable sound velocity, reverberation, noise, and echoes, while optical stereo suffers turbidity, blur, unstable illumination, and refraction.Improved channel models or cooperative opti-acoustic calibration can compensate these deficiencies and improve precision,.Source: 3) Three-dimensional underwater video… · pp. 24–25

C. Distributed and Cloud-based BMD Processing

  • Apache is the dominant distributed big-data analytics platform, appearing in about 95% of reported scientific articles, while cloud-based processing is increasing from 58% of businesses in 2017 to 73% in 2018,.The surveyed platforms provide distributed processing frameworks or cloud computing services, and the paper describes an expected shift toward cloud services,,.Source: C. Distributed and Cloud-based BMD… · p. 25, Figure 8 · p. 25
  • IoUT BMD can use edge, distributed, or cloud processing with SaaS, PaaS, and IaaS models; PaaS balances convenience and cost, while IaaS is cost-effective and includes provider-maintained hardware.Source: C. Distributed and Cloud-based BMD… · pp. 25–26
  • Spark runs workloads about 100× faster than Hadoop, supports multiple languages, complex machine-learning processing, and batch or streaming data, but requires high memory.The four Apache frameworks discussed are Spark, Hadoop, Storm, and Flink, all presented as suitable for high-speed IoUT BMD analytics.Source: C. Distributed and Cloud-based BMD… · pp. 25–26, Table X · p. 26
  • Storm reliably processes unbounded streams in real time but has high processing latency, low encryption capabilities, and high resource consumption, whereas Flink parallelizes large-scale applications into concurrently executed cluster tasks.Flink also supports real-time processing of pipelined streaming data.Source: C. Distributed and Cloud-based BMD… · p. 26

D. Marine Data Applications · 1) Coastal monitoring and GIS data:

IoUT sensor data supports marine applications by deriving geographic information on vessel locations, ocean conditions, species, and physical environments. Coastal monitoring and GIS infrastructures extend these capabilities through maritime databases, ecosystem monitoring, biogeographical observations, and georeferenced marine locations.

D. Marine Data Applications

  • IoUT sensor data enables applications including maritime operations, underwater localization, and marine-life tracking, with benefits for businesses using their own data.Source: D. Marine Data Applications · p. 26

1) Coastal monitoring and GIS data:

  • IoUT Geographic Data combines marine-vehicle localization, location-specific weather and climate information, and underwater-species recognition, counting, and distribution.These parameters describe physical environments, inhabitants, locations, and objects within particular sea areas.Source: 1) Coastal monitoring and GIS data · p. 26
  • Maritime GIS databases are essential for efficient international transportation and can integrate with IoUT infrastructures to support tracking, routing, and ecosystem monitoring.Researchers can use these merged systems for up-to-date monitoring of global ecosystems.Source: 1) Coastal monitoring and GIS data · pp. 26–27
  • Kalyvas et al. surveyed more than 180 open-access GIS databases and organized them into 19 classes, summarized here as tracking, cartographic, climate, and commerce data.Source: 1) Coastal monitoring and GIS data · p. 27
  • Biogeographical data extends the GIS categories by mapping underwater-species distributions through UNESCO-funded OBIS, which connects over 500 databases in 56 countries and records 120,000 species to 11,000 m depth.Source: 1) Coastal monitoring and GIS data · p. 27
  • Georeferenced locations fill a missing marine-cartographic subcategory by associating physical-map points with precise locations through Marine Regions, which lists 55,000 marine places.The list draws on the VLIMAR Gazetteer and MARBOUND databases.Source: 1) Coastal monitoring and GIS data · p. 27
  • The GIS discussion leads to underwater-positioning applications requiring precise undersea locations, including localization of sensors and vehicles and tracking of marine life.Source: 1) Coastal monitoring and GIS data · p. 27

2) Underwater localization: … A. Machine Learning Techniques

The paper surveys underwater localization and marine-life tracking within IoUT, then frames their outputs as Big Marine Data requiring distributed processing and ML/DL methods tailored to large, heterogeneous data.

2) Underwater localization:

  • Underwater localization is difficult because GPS signals do not penetrate seawater, while latency, fading, Doppler shifts, sparse nodes, and mobility prevent any single flawless navigation technique.IoUT systems therefore combine blind positioning, acoustic transponders, range sonars, image-based positioning, and SLAM.Source: 2) Underwater localization · p. 27
  • Blind positioning combines inertial and environmental sensors for autonomous localization, but accumulated errors can become unbounded; integrated sensing improves estimates while retaining power efficiency.Typical sensors include compasses, pressure sensors, Doppler velocity logs, gyroscopes, and accelerometers.Source: 2) Underwater localization · p. 27
  • Acoustic localization measures signal time of flight and acoustic speed, using arrays, fixed beacons, or cooperative modems to estimate positions relative to ships, buoys, or other vehicles.A single fixed beacon reduces location uncertainty and is cheaper and easier to install than multi-beacon long-baseline systems.Source: 2) Underwater localization · pp. 27–28
  • Image-based positioning matches successive seabed images for visual odometry, while stereoscopic cameras capture 3D transformations; SLAM combines mapping and localization to limit accumulated error.Image-based methods require substantial processing for feature extraction, and SLAM can use cameras, sonars, or inertial sensors, [171],,,.Source: 2) Underwater localization · p. 28

3) Marine life tracking:

  • Marine-life tracking requires specialized methods because animals may be too small to carry inertial sensors, transponders, sonars, or cameras, and GPS signals penetrate seawater only a few meters.Data-storage tags can archive measurements for later retrieval.Source: 2) Underwater localization · p. 28
  • Tracking options include RFID, acoustic tags, and image-based systems, with acoustic tags identifying animals through embedded IDs and hydrophone arrays, while image systems recognize behavior but must process many images promptly.RFID reaches about 10 m in freshwater but performs poorly in seawater; image tracking can count, measure, and study animals,.Source: 3) Marine life tracking · p. 28, 2) Underwater localization · p. 28

E. Section Summary

  • The BMD section organizes IoUT data acquisition, aggregation, fusion, signal/image/video types, observatories, underwater distortions, distributed and cloud processing, GIS, localization, and marine-life tracking.These sensory, imagery, GIS, tracking, and geo-tagging methods generate BMD that must subsequently be cleansed, processed, and analyzed with ML.Source: E. Section Summary · p. 29

IV. MACHINE LEARNING FOR BMD ANALYTICS

  • ML for BMD includes supervised and unsupervised learning, generative and discriminative models, conventional analytics, traditional NNs, and deep NNs; discriminative and deep approaches suit high-dimensional IoUT data.Discriminative models are generally simpler, faster, and lower-parameter, whereas deeper networks extract features but require more training data; the survey evaluates IoUT-suitable architectures and use cases, [177], [178].Source: §IV · p. 29, A. Machine Learning Techniques · p. 29

B. Deep Learning Frameworks and Libraries · C. Marine Data Cleaning for Machine Learning · 1) Underwater sensor data cleaning:

The paper presents open-source deep-learning frameworks for processing marine data and emphasizes that underwater BMD requires extensive preprocessing because harsh environmental conditions compromise data quality. It then introduces sensor-data cleaning around missing values, contaminated measurements, and outlier detection.

B. Deep Learning Frameworks and Libraries

  • Open-source deep-learning frameworks and libraries support the development of architectures for different marine-data applications.Table XII compares a range of these tools and frameworks.Source: B. Deep Learning Frameworks and… · p. 30, Table XII · p. 32
  • All listed frameworks can operate on NVIDIA CUDA-supported GPUs using parallel processing, whereas only some support OpenMP shared-memory multiprocessing.Shared-memory methods are identified as important for accelerating deep-network training.Source: B. Deep Learning Frameworks and… · p. 30
  • After data collection, preprocessing is required to ensure that training data are clean and do not create later learning problems.The paper transitions from marine-data architectures to data-cleaning procedures before discussing machine-learning processing in more detail.Source: B. Deep Learning Frameworks and… · p. 30

C. Marine Data Cleaning for Machine Learning

  • Underwater BMD preprocessing is crucial because environmental factors and hostile conditions make clean data difficult to acquire and threaten data sources.Data preparation and cleansing commonly consume over half of an ML project’s duration.Source: C. Marine Data Cleaning for Machine… · p. 30

1) Underwater sensor data cleaning:

  • Sensors are essential across IoUT subsystems and underwater exploration and surveillance applications, continuously generating large volumes of physical, chemical, and biological measurements.Examples of sensor types and data are listed in Table II.Source: 1) Underwater sensor data cleaning · p. 30
  • Underwater sensory-data cleansing addresses missing values, contaminated measurements, and outlier detection.The paper presents several alternative techniques for handling missing sensor values.Source: 1) Underwater sensor data cleaning · p. 30
  • Missing sensor values can be handled by deleting the record, inserting a global constant, or using statistical values such as the mean or median.Other listed options include inter-class statistical values and the most-likelihood value.Source: 1) Underwater sensor data cleaning · p. 30
  • Additional missing-value strategies use inter-class statistical values or the most-likelihood value.These alternatives supplement record deletion, global constants, and ordinary statistical imputation.Source: 1) Underwater sensor data cleaning · p. 30

MXNET ASF … D. Feature Extraction for Marine Data Analytics

The section surveys software and preprocessing methods for IoUT marine data, covering sensory-data denoising and outlier detection, underwater image enhancement and binning, and feature extraction with dimensionality reduction. It emphasizes tailoring descriptors and cleansing methods to difficult underwater conditions while supporting efficient machine-learning processing.

MXNET ASF

  • Underwater sensory-data cleansing addresses noise through SNR-guided method selection and detects outliers using IQR, T2, and Q criteria or advanced statistical approaches.Candidate denoising methods include low-pass filtering, regression, data binning, and wavelets; low-pass filters are most common, whereas wavelet methods are most complex to implement.Source: MXNET ASF · pp. 32–33

2) Underwater image and video data cleaning and quality enhancement:

  • Underwater image and video quality is degraded by absorption, scattering, chromatic aberrations –, and light-source noise, requiring cleansing and enhancement.Automatic quality assessment combines colorfulness, contrast, and sharpness into an objective metric used to guide enhancement and bandwidth-related decisions,,.Source: 2) Underwater image and video data… · p. 33
  • Software-based image preparation is generally cheaper than bespoke imaging hardware and can enhance underwater images and videos after quality assessment.Ancuti et al.’s camera-independent method jointly dehazes and enhances images by improving contrast, sharpening edges, and exposing dark regions without prior environmental knowledge.Source: 2) Underwater image and video data… · p. 33

3) Underwater image data binning:

  • Image binning groups similar pixels into partitions to reduce noise and data complexity, and is recommended before feature extraction to reduce computational cost and accelerate machine learning.Each pixel receives a label identifying its parent partition through a membership index.Source: 3) Underwater image data binning · p. 33
  • No image-partitioning method is universally accepted, so selecting an appropriate clustering approach remains challenging for a given underwater application.Methods may exploit discontinuities such as edges or similar regions; Table XIII lists clustering methods beneficial to underwater applications.Source: 3) Underwater image data binning · p. 33, Table XIII · p. 34

D. Feature Extraction for Marine Data Analytics

  • Feature extraction replaces raw sensory vectors and image matrices with low-dimensional numerical descriptors, but descriptor design is intricate and deep networks remain difficult to train with limited datasets.Underwater conditions make feature discovery harder because of non-uniform illumination, variable brightness, and degraded colors.Source: D. Feature Extraction for Marine Data… · p. 33
  • Underwater descriptors span color, texture, shape, and other classes, with application-specific combinations reaching 66 features in Huang et al.’s Fish Recognition project.After feature selection, correlation analysis and dimensionality reduction retain independent descriptors using reduction, selection, or aggregation techniques.Source: D. Feature Extraction for Marine Data… · pp. 33–34
  • ML- and DL-aided IoUT software components also rely on appropriate high-performance hardware platforms.Source: D. Feature Extraction for Marine Data… · p. 34

E. Hardware Platforms for ML in IoUT · F. ML Techniques in BMD Applications · 1) Deep NNs for static IoUT data:

The survey links IoUT machine-learning deployment to parallel hardware architectures and reviews deep neural networks for static marine data, emphasizing their processing roles and implementation trade-offs. It identifies GPUs, embedded Jetson systems, ASICs, and FPGAs as practical platforms, while characterizing autoencoders, DBNs, RNTNs, and CNNs for static data analytics.

E. Hardware Platforms for ML in IoUT

  • ML processing hardware is organized into shared-memory multiprocessors and networked processor systems, with parallelism increasing overall throughput.Shared-memory platforms include ASICs, FPGAs, multi-core CPUs, and GPUs, while distributed systems coordinate processors through message passing,.Source: E. Hardware Platforms for ML in IoUT · p. 34
  • Distributed processor systems parallelize ML through data, model, or pipelined parallelization, respectively dividing datasets, algorithm components, or sequential processing stages.Data parallelization exchanges parameter estimates and includes MapReduce,; model parallelization exchanges partial error vectors [234],; pipelining overlaps records across chained processors.Source: E. Hardware Platforms for ML in IoUT · pp. 34–35
  • GPUs offer affordable resources and high speed for scientific and industrial underwater applications, including obstacle detection, SLAM localization, and coral detection,.CUDA also supports general-purpose GPU processing, including real-time underwater acoustic beamforming with fast Fourier transforms.Source: E. Hardware Platforms for ML in IoUT · pp. 35–36
  • Embedded Jetson systems provide compact, high-performance, low-power GPU computing that enables low-latency deep-learning inference in underwater vehicles and platforms.Their Linux Ubuntu environment supports frameworks such as TensorFlow, while their small form factor addresses conventional GPUs’ dependence on bulky host computers.Source: E. Hardware Platforms for ML in IoUT · pp. 35–36
  • ASICs and FPGAs combine low form factor, high throughput, and power efficiency, supporting industry-scale and real-time low-power IoUT applications.A standalone FPGA architecture for underwater-sonar noise suppression was reported to overcome sonar-based underwater-visibility problems.Source: E. Hardware Platforms for ML in IoUT · p. 36

1) Deep NNs for static IoUT data:

  • For static IoUT data, fully connected networks, DBNs, CNNs, autoencoders, and RNTNs are presented as suitable deep-learning architectures.RNTNs process segmented static underwater images to determine hierarchical categories such as background, coral, plants, and fish.Source: 1) Deep NNs for static IoUT data · p. 36
  • Autoencoders support unsupervised clustering, feature extraction, dimensionality reduction, compression, decompression, and noise reduction through separate or combined encoder-decoder components.The encoder extracts features, while the decoder reconstructs the input; using both simultaneously produces a noise-reducing neural network.Source: 1) Deep NNs for static IoUT data · p. 36
  • DBNs use consecutive restricted Boltzmann machine pairs to reduce feature dimensions and can perform unsupervised clustering or supervised classification depending on termination and training labels.A mid-layer can serve as a feature-extraction output, whereas reducing the final layer to the class count enables labeled classification.Source: 1) Deep NNs for static IoUT data · p. 36

2) Deep NNs for dynamic IoUT data: · 3) ML solutions in underwater image applications: · 4) ML solutions in underwater video applications:

The surveyed methods use RNNs for nonlinear dynamic IoUT data and apply deep learning across underwater image and video tasks, including enhancement, recognition, tracking, mosaicing, and marine-life analysis. Evidence indicates that appropriately designed DNNs can outperform traditional models in sonar recognition, while cascaded classifiers improve precision at added hardware cost.

2) Deep NNs for dynamic IoUT data:

  • RNNs and variants such as LSTM and GRU are the deep NN architectures identified for nonlinear dynamic systems, using time-series and sequential streams to construct supervised classifiers.The surveyed dynamic-data examples include continuous underwater sensor outputs.Source: 2) Deep NNs for dynamic IoUT data · p. 37

3) ML solutions in underwater image applications:

  • Underwater image applications use deep learning to enhance images degraded by noise, absorption, scattering, and color distortion, while image-quality assessment and clustering remain developing areas.Novel deep learning methods have not yet received sufficient attention for underwater image clustering, although embedded clustering can support unsupervised segmentation.Source: 3) ML solutions in underwater image… · p. 37
  • Cascading classifiers can improve underwater recognition precision by combining shape- and texture-based classifiers or object-specific binary SVM detectors.Object-specific detectors support high-precision multi-object recognition, but combining them requires more hardware resources than a single classifier.Source: 3) ML solutions in underwater image… · p. 37
  • Appropriately designed and well-trained DNNs may outperform statistical and traditional models for underwater sonar object recognition, while CNNs avoid HOG-based feature extraction preprocessing.The comparison uses Average Precision and includes deep NNs, statistical methods, and traditional NNs.Source: 3) ML solutions in underwater image… · p. 38, Table XVIII · p. 38

4) ML solutions in underwater video applications:

  • Machine learning supports underwater video tracking in visible-light and sonar imagery, with deep NNs handling moving-object localization and detection, including in turbid environments.Sonar provides long-range, low-data-rate imaging where visible-light video is limited.Source: 4) ML solutions in underwater video… · p. 38
  • Machine-learning photo mosaicing combines visible-light or sonar frames for exploration and navigation, and supports submersible localization and positioning through SLAM.The passage describes this capability as routinely performed by state-of-the-art underwater vehicles.Source: 4) ML solutions in underwater video… · p. 38
  • Machine-learning video processing enables marine-life studies to analyze high-volume observatory footage for behavior understanding, abundance counting, size measurement, detection, recognition, and tracking.These applications help marine biologists extract desired information from large underwater video collections.Source: 4) ML solutions in underwater video… · pp. 38–39

5) ML solutions in underwater sensor applications: … C. Development of Low-cost and Affordable Sensors

The surveyed IoUT and BMD landscape combines specialized machine-learning methods with major challenges in network management, energy autonomy, and sensor affordability. The paper identifies practical protocols, harvesting strategies, and cost-reduction approaches while emphasizing that underwater conditions require dedicated solutions.

5) ML solutions in underwater sensor applications:

  • RNN applications in IoUT remain rare, although forecasting, fault diagnosis, and underwater movement control are reported, with LSTM and GRU improving sensor-data prediction over traditional methods.Sea-temperature and salinity prediction is especially relevant because these variables affect acoustic communication and ecosystems.Source: 5) ML solutions in underwater sensor… · p. 39

G. Section Summary

  • The ML review spans statistical methods, traditional and deep neural networks, data cleaning, feature extraction, software frameworks, hardware platforms, and reported underwater precision and accuracy.Lightweight traditional neural networks may remain appropriate for limited-complexity underwater applications despite advances in sophisticated deep networks.Source: G. Section Summary · p. 39

V. CHALLENGES AND FUTURE DIRECTIONS IN IOUT AND BMD

  • IoUT advances are hindered because terrestrial technologies often perform poorly underwater, requiring coordinated research from analytics, engineering, hardware, and policy communities.The paper frames subsequent directions around challenges and opportunities in IoUT, BMD, and ML.Source: §V · p. 39

A. Underwater Network Management System

  • U-NMS must automatically monitor faults and allocate resources across heterogeneous underwater networks, extending IoT management through FCAPSC constraint management for mobility, fragility, dynamics, and hostile channels.U-SNMP covers network functionalities, while LWM2M supports device operations with adjustments; both have limitations, and OVSDB-based underwater SDN management requires further security research.Source: A. Underwater Network Management System · pp. 39–40

B. Energy Conservation and Harvesting in IoUT Devices

  • IoUT energy strategies include solar, tidal-wave, wireless, and wired transfer, rechargeable batteries, and energy-aware consumption, but most harvesting methods are unpredictable and wired infrastructure is costly.Energy-aware designs can extend network lifetime and improve QoS, while batteries add maintenance and system costs.Source: B. Energy Conservation and Harvesting… · pp. 40–41

C. Development of Low-cost and Affordable Sensors

  • Underwater sensors are expensive, motivating quality-cost trade-offs, inland sample assessment, inferential measurements, interdisciplinary design, miniaturization, and tailored low-cost sensing technologies.Electrical resistance can estimate dissolved ions more cheaply than electrochemical sensors, while optical fibers, miniature robots, and nano-sensors support lower-cost measurements.Source: C. Development of Low-cost and… · p. 41
  • Reported affordable sensors include a salinity-measuring patch antenna, a temperature-compensated fiber-Bragg-grating pressure sensor, a scattered-light turbidity sensor, and a low-frequency optical-backscatter fluorometer.These examples use inexpensive components or sensing principles to reduce underwater measurement costs while retaining targeted functionality.Source: C. Development of Low-cost and… · p. 41
  • Edge computing can support an onboard laboratory for preliminary sample analysis and remote video-enabled control, reducing the expense and time of inland evaluation or in-situ expert presence.The onboard lab transmits numerical results through the web after conducting limited-capability experiments on collected samples.Source: C. Development of Low-cost and… · p. 41

D. Large-scale IoUT Underwater Communications … H. Sparse and High-maintenance Sensing Devices in IoUT

IoUT faces severe communication, routing, security, modeling, and maintenance challenges because underwater links are short-range or bandwidth-limited, unreliable, energy-constrained, and exposed to harsh environments. The survey therefore emphasizes heterogeneous and edge-assisted communications, learning-based control and security, precise channel models, and self-managing sensing devices.

D. Large-scale IoUT Underwater Communications

  • Underwater electromagnetic and optical links are short-range, while acoustic links support longer ranges but suffer narrow bandwidth and cross-talk, limiting large-scale IoUT communications.Fiber-optic deployment and maintenance are also expensive underwater.Source: D. Large-scale IoUT Underwater… · p. 41
  • Heterogeneous gateways, satellite or aerial connections, and edge processing can extend IoUT reach and reduce transmitted data volume, but energy harvesting and underwater energy scarcity remain obstacles.MEC with autonomous underwater or aerial vehicles can process latency-sensitive data locally while mitigating big-data collection.Source: D. Large-scale IoUT Underwater… · p. 42

E. Dynamic IoUT Signal Routing and Traffic Control

  • Low channel capacity and concurrent transmissions make traffic control difficult, motivating energy-aware routing, low-complexity transceivers, deep learning, and SDN-based congestion management,,,.Routing must balance more short hops against fewer long hops while considering battery charge.Source: E. Dynamic IoUT Signal Routing and… · p. 42
  • For wide-scale IoUT, replacing whole-path actions with next-hop decisions and cascading CNN prediction with DBN action selection can reduce packet loss and improve throughput over conventional routing.The approach addresses exponential path growth that can cause a DNN to fail as network scale increases.Source: E. Dynamic IoUT Signal Routing and… · p. 42

F. Securing Underwater Networks

  • IoUT security is difficult because unreliable, delayed, energy-limited channels expose endpoints across layers, while conventional application-layer protocols are computationally power-intensive and no comprehensive energy-effective framework is available.Research should distribute security functions across layers and collect IoUT attack datasets for ML-based protection.Source: F. Securing Underwater Networks · p. 43
  • When attack datasets are unavailable, reinforcement learning can train during experiments; Xiao et al. learned to avoid jammed locations and adjust transmit power to achieve the required bit error rate.CNNs and RNNs are associated with single-attack detection, whereas DBNs may suit collections of attacks, but IoUT application requires further investigation.Source: F. Securing Underwater Networks · p. 43

G. Deleterious Effects of Imprecise Channel Modeling

  • Approximated channel models produce inaccurate communication and simulations, whereas precise models are needed for reliable preamble detection, transmission-redundancy optimization, and reinforcement-learning-based adaptive transmission.Future models should be both precise and computationally efficient.Source: G. Deleterious Effects of Imprecise… · pp. 43–44

H. Sparse and High-maintenance Sensing Devices in IoUT

  • Sparse IoUT sensing devices face costly maintenance from erosion, corrosion, sediments, pollution, and other seawater effects, motivating ML-based self-management with self-evaluation, self-configuration, and automatic reporting,.Such intelligent hardware and software could support autonomous decision-making.Source: H. Sparse and High-maintenance Sensing… · p. 44

I. Poor Underwater Positioning and Navigation … VI. CONCLUSION

The surveyed IoUT research highlights unresolved challenges in underwater positioning, nondestructive testing, data leveraging, deep-network training, and image quality, while reviewing communication, BMD, and ML/DL advances. The conclusion identifies open research directions spanning underwater communications, BMD processing, machine intelligence, and future IoUT applications.

I. Poor Underwater Positioning and Navigation

  • GPS cannot penetrate seawater, and standalone navigation methods accumulate positioning error; future systems should combine techniques while balancing accuracy, cost, and complexity.Examples include combining image-based positioning with SLAM and blind positioning with long-baseline acoustic transponders. Environmental surveys, deep networks, gravity-aided navigation, and LPWAN adaptation remain research opportunities.Source: §I · pp. 44–45

J. Non-Destructive Testing in Underwater Applications

  • Underwater NDT uses optical fibers and other equipment, including cameras, electromagnetic flux detectors, ultrasonic transceivers, and magnetic inductors, during normal DUT operation for pipeline, riser, downhole-tool, and water-resource applications –,.Source: J. Non-Destructive Testing in… · p. 45

K. Lack of Strong Data Leveraging Tools

  • BMD processing lacks strong automatic data-leveraging tools, especially for annotation and feature selection, because single features are usually inadequate for underwater classification and clustering.Deep models may help, but classical ML can offer fewer parameters, faster convergence, interpretability, and easier debugging; feature selection directly affects ML performance.Source: K. Lack of Strong Data Leveraging Tools · p. 45

L. Training Deep Networks

  • Deep networks automatically extract features but require substantial training data, creating a trade-off among data volume, network size, and performance.Proposed directions include data augmentation, GANs,, expert-generated data, architectural modification, and pruning for mobile devices.Source: L. Training Deep Networks · pp. 45–46

M. Degraded Underwater Images

  • Underwater photography remains degraded by environmental factors, motivating real-time elimination of forward- and back-scattering and improved hardware, software, and deep-learning reconstruction methods,,.Suggested hardware includes light polarization, multispectral imaging, and stereoscopic imaging; software includes wavelength compensation and color reconstruction,.Source: M. Degraded Underwater Images · p. 46

VI. CONCLUSION

  • The survey connects BMD processing with ML and DL, critically appraises hardware and software research, and identifies open issues, future directions, and prospective IoUT applications.The paper frames IoUT as a source of new directions in undersea data acquisition, communication, BMD handling, and oceanic data processing.Source: §VI · p. 46
  • The survey reviews IoUT communication technologies, channel modeling, simulation tools, topology and routing optimization, security, protocols, SDN, edge computing, and a five-layer TCP/IP architecture.Acoustic, electromagnetic, and optical communication models are considered under harsh underwater propagation conditions.Source: §VI · p. 46
  • IoUT generates BMD whose storage, transport, preparation, and analysis require solutions spanning sensors, imagery, video, geographic data, localization, databases, distributed processing, and cloud services.Manual and semi-automatic processing is no longer appropriate for the IoUT era.Source: §VI · p. 46

Related papers

Loading 2012.06712v2…