Machine Learning for the Internet of Underwater Things: From Fundamentals to Implementation

Kenechi Omeke; Attai Abubakar; Michael Mollel; Lei Zhang; Qammer H. Abbasi; Muhammad Ali Imran

eess.SYMar 8, 2026

Machine Learning for the Internet of Underwater Things: From Fundamentals to Implementation

Authors: Kenechi Omeke, Attai Abubakar, Michael Mollel, Lei Zhang, Qammer H. Abbasi, Muhammad Ali Imran

Organizations: James Watt School of Engineering, University of Glasgow, Glasgow, United Kingdom

Abstract

The Internet of Underwater Things (IoUT) is becoming a critical infrastructure for ocean observation, marine resource management, and climate science. Its development is hindered by severe acoustic attenuation, propagation delays far exceeding those of terrestrial wireless systems, strict energy constraints, and dynamic topologies shaped by ocean currents. Machine learning (ML) has emerged as a key enabler for addressing these limitations, offering data driven mechanisms that enhance performance across all layers of underwater wireless sensor networks. This tutorial survey synthesises ML methodologies supervised, unsupervised, reinforcement, and deep learning specifically contextualised for underwater communication environments. It outlines the algorithmic principles of each paradigm and examines the conditions under which particular approaches deliver superior performance. A layer wise analysis highlights physical layer gains in localisation and channel estimation, MAC layer adaptations that improve channel utilisation, network layer routing strategies that extend operational lifetime, and transport layer mechanisms capable of reducing packet loss by up to 91 percent. At the application layer, ML enables substantial data compression and object detection accuracies reaching 92 percent. Drawing on 300 studies from 2012 to 2025, the survey documents energy efficiency gains of 7 to 29 times, throughput improvements over traditional protocols, and cross layer optimisation benefits of up to 42 percent. It also identifies persistent barriers, including limited datasets, computational constraints, and the gap between theoretical models and real world deployment. The survey concludes with emerging research directions and a technology roadmap supporting ML adoption in operational underwater networks.

Figures & tables

Panel of Figure 1 — Fig. 1: Taxonomy of ML techniques for the IoUT. This survey covers four major categories: supervised learning for classification and prediction tasks, unsupervised learning for pattern discovery and compression, reinforcement learning for adaptive protocol design, and advanced paradigms including federated learning and transformer architectures.

Function	Traditional Approach	ML-Based Approach
Channel Estimation	Analytical models, pilot symbols	Neural network prediction, adaptive [ 34 ]
Power Control	Fixed levels, lookup tables	RL-based adaptation [ 35 ]
MAC Protocol	Fixed backoff, TDMA slots	Learning-based scheduling [ 43 ]
Routing	Shortest path, geographic	Q-learning, GNN-based [ 44 ]
Localisation	ToA/TDoA algorithms	DNN regression, RL-aided [ 55 ]

TABLE I: Comparison of Traditional vs. ML-Based Approaches in Underwater Communications

Stage	Layer Type	Output/Kernel
Encoder	Conv $\to$ ReLU	64 filters, $5\times 5$
	MaxPool	$2\times 2$
	Conv $\to$ ReLU	32 filters, $3\times 3$
	MaxPool	$2\times 2$
	Dense	16 units
Decoder	Dense	32 units

TABLE II: Autoencoder Architecture for Underwater Acoustic Compression

Stream	Layer / Action	Parameters
Sonar	Conv2D $\times$ 3	32, 64, 64 filters
	Kernels	$8\times 8$ , $4\times 4$ , $3\times 3$
	Strides	4, 2, 1
Sensor	Dense $\times$ 2	128, 64 units
Joint	Concatenate	Fusion of streams
	Dense	512 units

TABLE III: DQN Architecture for AUV Navigation

Layer Type	Filters	Kernel / Pool
Conv3D + BN	32	$5\times 5\times 3$
Pool3D	-	$2\times 2\times 1$
Conv3D + BN	64	$3\times 3\times 3$
Pool3D	-	$2\times 2\times 2$
Conv3D + BN	128	$3\times 3\times 3$
Pool3D	-	$2\times 2\times 2$

TABLE IV: 3D CNN Architecture for Sonar Classification

Category	Technique	Primary Applications	Key Advantages
Supervised Learning	k-NN, SVM	Modulation classification, vessel identification	Robust with limited data [ 42 ]
	Random Forests	Fault detection, routing decisions	Interpretable, handles mixed features [ 74 ]
	Gaussian Processes	Field estimation, path planning	Uncertainty quantification [ 56 ]
Unsupervised Learning	k-Means, DBSCAN	Network clustering, topology organisation	Adapts to irregular deployments [ 78 ]
	PCA, Autoencoders	Data compression, anomaly detection	50–1000 $\times$ compression [ 64 ]
Reinforcement Learning	Q-Learning, DQN	MAC protocols, power control, routing	Learns from interaction [ 44 ]

TABLE V: Summary of ML Techniques for Underwater Communications

Panel of Figure 7 — Fig. 2: Taxonomy of existing surveys related to ML and IoUT. Our survey uniquely integrates knowledge across all four categories, providing comprehensive coverage that individual surveys lack.

ML in Wireless Sensor Networks (General)
Survey Reference	Year	Primary Scope	Key Contributions	Limitations
Alsheikh et al. [ 65 ]	2014	ML algorithms in terrestrial WSNs	Comprehensive ML taxonomy, algorithm comparison, complexity analysis	No underwater considerations, outdated ML techniques
Kumar et al. [ 66 ]	2019	ML techniques for WSN optimisation	Energy efficiency focus, clustering algorithms, fault detection	Minimal underwater content, lacks deep learning
Klaine et al. [ 136 ]	2017	Self-organising networks with ML	SON paradigm, cellular focus, optimisation techniques	Terrestrial only, no acoustic channels
Underwater Communications and Networking
Li et al. [ 12 ]	2019	High reliability in UWSNs	Protocol comparison, reliability metrics, cross-layer design	No ML techniques discussed

TABLE VI: Comprehensive Comparison of ML and Underwater Networking Surveys (2012–2025)

Panel of Figure 9 — Fig. 3: Coverage matrix comparing existing surveys across protocol layers (Physical, MAC, Network, Transport, Application), cross-layer optimisation, ML paradigms (DL: Deep Learning, RL: Reinforcement Learning, FL: Federated Learning), and implementation guidance. Our survey provides comprehensive coverage across all dimensions.

Panel of Figure 10 — Fig. 4: Four critical gaps identified in existing survey literature and the corresponding solutions provided by this survey. Each gap represents a significant barrier to ML adoption in IoUT systems that our comprehensive treatment addresses.

Physical Layer
Application	Baseline	ML Method	Improvement
Localisation	Trilateration	CNN	7 $\times$ accuracy
Channel Est.	LS Pilot	LSTM	15 dB MSE gain
Modulation Class.	Energy Det.	ResNet	25% @ -5 dB SNR
MAC Layer
Channel Access	ALOHA	Q-learning	45% throughput

TABLE VII: Sample Quantitative Comparisons Provided in This Survey

Panel of Figure 12 — Fig. 5: Five unique value propositions distinguishing this survey from existing literature, providing comprehensive coverage from tutorial foundations through future research directions.

Task	Method	Metric	Value	Ref.
Localisation	Trilateration	Error	8.5 m	[ 145 ]
	$k$ -NN	Error	1.2 m	[ 120 ]
	CNN	Error	0.8 m	[ 55 ]
Channel Est.	LS Pilot	MSE	-5 dB	[ 22 ]
	CNN	MSE	-15 dB	[ 34 ]
	LSTM	MSE	-20 dB	[ 76 ]

TABLE VIII: Sample from Quantitative Performance Repository

Panel of Figure 14 — Fig. 6: Three reading patterns supported by our organisation: vertical for layer-specific expertise, horizontal for technique comparison, and diagonal for algorithm tracing.

Algorithm	BER @ 0 dB	Train	Infer.	Mem.
MMSE Equalizer	$3.2\times 10^{-2}$	N/A	0.5 ms	10 KB
RLS Adaptive	$1.8\times 10^{-2}$	N/A	2.0 ms	25 KB
MLP (3 layer)	$8.4\times 10^{-3}$	2 h	5.0 ms	150 KB
CNN (5 layer)	$4.2\times 10^{-3}$	8 h	12 ms	500 KB
LSTM	$2.1\times 10^{-3}$	24 h	20 ms	1.2 MB
Transformer	$1.3\times 10^{-3}$	48 h	35 ms	4.5 MB

TABLE IX: Integrated Benchmark: Acoustic Channel Equalisation

Panel of Figure 16 — Fig. 7: Positioning of this survey relative to existing works across ML depth (from basic algorithms to advanced paradigms) and IoUT breadth (from single-application to full protocol stack). Our survey uniquely occupies the ideal region combining comprehensive IoUT coverage with advanced ML treatment.

Feature	This	Jan.	Khi.	Chr.
Full protocol stack	✓	–	$\sim$	–
Deep learning coverage	✓	$\sim$	–	$\sim$
Reinforcement learning	✓	–	$\sim$	✓
Federated learning	✓	–	–	–
Cross-layer analysis	✓	–	–	–
Quantitative repository	✓	$\sim$	$\sim$	$\sim$

TABLE X: Final Comparison: This Survey vs. Closest Existing Works

Panel of Figure 18 — Fig. 8: Mapping of ML techniques to IoUT protocol stack layers. Each layer employs specialised algorithms suited to its unique challenges, while cross-layer optimisation enables holistic system improvement.

Method	Mean Err.	95% Err.	Latency	Robustness
Trilateration	8.5 m	22 m	10 ms	Poor
Weighted LS	6.2 m	18 m	25 ms	Fair
Particle Filter	3.8 m	11 m	200 ms	Good
$k$ -NN ( $k$ =10)	1.2 m	3.5 m	15 ms	Excellent
CNN	0.8 m	2.2 m	50 ms	Excellent
DQN Active	0.5 m	1.5 m	100 ms	Excellent

TABLE XI: Localisation Performance Comparison in 100m

\times

100m

\times

50m Volume

Application	Metric	Traditional	ML-Based	Improvement	ML Technique
Physical Layer
Localisation	Position Error (m)	8.5	0.8	91% reduction	CNN
Channel Estimation	MSE	0.043	0.012	Significant reduction	LSTM
Modulation Classification	Accuracy @ 0 dB	75%	96%	28% increase	CNN
Adaptive Modulation	Throughput (kbps)	Baseline	+20–45% ∗	Substantial increase	DQN
MAC Layer

TABLE XII: Comprehensive ML vs. Traditional Performance Comparison Across IoUT Applications. Values represent typical results from cited studies; actual performance varies with experimental conditions and environments. See Section V-A and cited papers for detailed methodology and contexts.

Operation	Traditional	ML-Based	Savings
Acoustic Transmission	10 J/pkt	0.34 J/pkt	29 $\times$
Channel Estimation	0.5 J/est	0.08 J/est	6 $\times$
Route Discovery	45 J/route	2.1 J/route	21 $\times$
Object Detection	8.2 J/frame	0.15 J/frame	55 $\times$
Network Maintenance	850 J/day	12 J/day	71 $\times$
Total Daily	2800 J	180 J	1556 $\times$

TABLE XIII: Energy Efficiency Gains: ML vs. Traditional Approaches

Training/Setup Phase
Task	Traditional	ML-Based
Localisation Setup	$O(n^{3})$	$O(n^{2}d)$
Routing Table	$O(n^{3})$	$O(n^{2}kE)$
Channel Model	$O(T^{2})$	$O(TBE)$
Inference/Operation Phase
Position Estimation	$O(n^{2})$	$O(k)$

TABLE XIV: Computational Complexity: ML vs. Traditional Algorithms

Application Domain	Metric	Traditional	ML-Based	Improvement	ML Technique	Reference
Physical Layer
Localisation	Position Error (m)	8.5	0.8	91% reduction	CNN	[ 168 ]
Channel Estimation	MSE	0.043	0.012	Significant reduction	LSTM	[ 151 ]
Modulation Classification	Accuracy @ 0dB SNR	75%	96%	28% increase	CNN	[ 169 ]
Adaptive Modulation	Throughput (kbps)	Baseline	Improved	Substantial increase	DQN	[ 170 ]
MAC Layer

TABLE XV: Comprehensive ML vs Traditional Performance Comparison Across IoUT Applications

Operation	Traditional	ML-Based	Savings	Ref.
Acoustic Transmission	10 J/packet	0.34 J/packet	29 $\times$	[ 186 ]
Channel Estimation	0.5 J/estimate	0.08 J/estimate	6 $\times$	[ 151 ]
Route Discovery	45 J/route	2.1 J/route	21 $\times$	[ 176 ]
Object Detection	8.2 J/frame	0.15 J/frame	55 $\times$	[ 179 ]
Network Maintenance	850 J/day	12 J/day	71 $\times$	[ 78 ]
Total Daily Energy	2800 J	180 J	15.6 $\times$	—

TABLE XVI: Energy Efficiency Gains: ML vs Traditional Approaches

Panel of Figure 25 — Fig. 9: Scalability comparison showing packet delivery ratio vs. network size. ML approaches maintain consistent performance while traditional protocols degrade significantly with scale.

Condition	Metric	ML Gain	Reference
High node mobility	PDR	+45%	[ 188 ]
Sparse topology	Delivery ratio	+38%	[ 29 ]
High traffic load	Throughput	+67%	[ 155 ]
Time-varying channel	BER	-52%	[ 170 ]
Low SNR ( $<$ 0 dB)	Classification	+28%	[ 169 ]
Multi-hop (5+ hops)	Latency	-41%	[ 164 ]

TABLE XVII: ML Performance Gains Under Varying Conditions

Training/Setup Phase
Task/Algorithm	Traditional	ML-Based
Localisation Setup	$O(n^{3})$	$O(n^{2}d)$
Routing Table Creation	$O(n^{3})$	$O(n^{2}kE)$
Channel Model Fitting	$O(T^{2})$	$O(TBE)$
Clustering Initialisation	$O(n^{2}\log n)$	$O(nkE)$
Inference/Operation Phase

TABLE XVIII: Computational Complexity: ML vs Traditional Algorithms

Algorithm	RAM	Flash	Inference	Platform
Q-Learning Router	12 KB	48 KB	0.3 ms	Cortex-M4
DQN Router	64 KB	256 KB	2.1 ms	Cortex-A53
LSTM Predictor	128 KB	512 KB	5.4 ms	Jetson Nano
CNN Classifier	256 KB	1.2 MB	8.7 ms	Coral TPU
Dijkstra (100 nodes)	40 KB	8 KB	12.3 ms	Cortex-M4
AODV (100 nodes)	120 KB	24 KB	45.7 ms	Cortex-M4

TABLE XIX: Practical Resource Requirements for Embedded Deployment

Panel of Figure 29 — Fig. 10: Pareto frontier for underwater object detection showing accuracy-energy trade-offs. Blue points represent Pareto-optimal configurations; red points are dominated solutions.

Operation	Traditional	ML	Hardware
Channel Estimation	450 mJ	72 mJ	ARM Cortex-M4
Route Computation	890 mJ	23 mJ	ARM Cortex-A53
Object Detection	8200 mJ	150 mJ	Jetson Nano
Anomaly Detection	340 mJ	45 mJ	Coral TPU
Packet Scheduling	125 mJ	18 mJ	ARM Cortex-M4
Cluster Formation	560 mJ	85 mJ	ARM Cortex-A53

TABLE XX: Energy Consumption per Operation

Protocol Comparison	Improvement	Network Size	Reference
QELAR vs VBF	20% longer	100 nodes	[ 176 ]
DEKCS vs LEACH	70% longer	200 nodes	[ 78 ]
EDORQ vs DBR	35% longer	150 nodes	[ 191 ]
Q-EAVAR vs QELAR	25% longer	100 nodes	[ 192 ]
ENCRQ vs QHUC	23.5% longer	200 nodes	[ 177 ]
CTRGWO vs LEACH	23.5% longer	150 nodes	[ 193 ]

TABLE XXI: Network Lifetime Improvements: ML vs Traditional Protocols

Performance Dimension	Typical Gain	Maximum Reported
Localisation accuracy	5–10 $\times$	99.98% accuracy
Throughput	1.5–2.5 $\times$	148% increase
Energy efficiency	6–70 $\times$	1556 $\times$
Network lifetime	1.5–3 $\times$	173% extension
Scalability (500 nodes)	3–4 $\times$ PDR	265% advantage
Inference latency	2–20 $\times$ faster	$O(1)$ vs $O(n^{2})$

TABLE XXII: Summary of ML Performance Advantages

Panel of Figure 33 — Fig. 11: Environmental adaptation framework for underwater ML systems. Red arrows indicate environmental challenges affecting system performance; green arrows represent adaptive solutions; the yellow output block shows the maintained performance through integrated adaptation strategies.

Deployment Scenario	Data Strategy	Update Strategy	Learning Strategy	Key Considerations
Short-term ( $<$ 1 month)	Pre-collected	None required	Pre-trained only	Minimise complexity
Medium-term (1–12 months)	Transfer learning	Differential updates	Online adaptation	Balance adaptability vs. stability
Long-term ( $>$ 1 year)	Active learning + synthetic	Hierarchical federated	Continual learning	Prevent catastrophic forgetting
Deep sea ( $>$ 1000m)	Synthetic + few-shot	Physical recovery only	Edge learning	Extreme isolation constraints
Mobile (AUV-based)	Opportunistic collection	Gossip-based	Collaborative learning	Exploit mobility for updates

TABLE XXIII: Deployment Strategy Selection Guide for Underwater ML Systems

Challenge Category	Specific Challenge	Recommended Solution	Expected Outcome
Resource Constraints	Limited processing	TinyML, quantisation, pruning	10–100 $\times$ speedup
	Memory limitations	Model compression, tiling	4–32 $\times$ reduction
	Energy budget	Adaptive duty cycling, hierarchical inference	85 $\times$ energy reduction
Environmental	Biofouling	Fouling-aware training, online calibration	85% accuracy at 3 months
	Corrosion	Predictive maintenance, redundancy	87% failure prediction
	Pressure effects	Depth-aware models	92% accuracy maintained

TABLE XXIV: Implementation Challenges and Solutions Summary

Technology	Near-term (1–2 years)	Mid-term (3–5 years)	TRL	Key Enablers
Physics-Informed NN	Broadband acoustic modelling, Single-domain PINNs	Multi-physics coupling, Real-time deployment	4–5	GPU acceleration, Automatic differentiation
Transformers	UATR classification, Signal denoising	End-to-end communication, Multi-modal fusion	5–6	Sparse attention, Model compression
Graph Neural Nets	Secure routing, Trust modelling	Dynamic topology learning, Distributed inference	4–5	Efficient message passing, Edge deployment
Federated Learning	Privacy-preserving training, Model aggregation	Asynchronous underwater FL, Byzantine robustness	4–5	Communication-efficient protocols
Meta-Learning	Few-shot environment adaptation	Continual meta-learning, Zero-shot deployment	3–4	Diverse training environments
Semantic Comm.	Task-specific compression, Meaning encoding	LLM-guided semantics, Generative decoding	3–4	Edge-optimised LLMs, Diffusion models

TABLE XXV: Research Roadmap for ML-Enabled Underwater Communications

Panel of Figure 37 — Fig. 12: Overview of emerging ML technologies and their integration pathways for intelligent underwater networks. Core ML technologies (top) enable system-level integration (middle) that supports next-generation underwater applications (bottom).

Challenge Category	Key Problem	Research Direction	Potential Impact
Data Scarcity	Million-dollar datasets	Self-supervised learning, PINNs	100 $\times$ reduction in data needs
Interpretability	Black-box decisions	Physics-grounded explanations	Enable regulatory approval
Real-Time Processing	Computational gap	TinyML, neuromorphic computing	100 $\times$ efficiency improvement
Adversarial Robustness	Natural/intentional attacks	Physics-constrained defences	Maintain 95%+ accuracy under attack
Federated Learning	Communication constraints	Hierarchical FL, compression	Enable collaborative training
Maintenance	Model drift in isolation	Continual learning, self-healing	Extend deployment 3–5 $\times$

TABLE XXVI: Summary of Research Priorities for ML in IoUT Systems

Panel of Figure 39 — Fig. 13: Taxonomy of open challenges in ML for IoUT systems. By aligning Federated Learning with Real-Time constraints, the vertical dependency is clarified. Bidirectional arrows (orange) show the interplay between technical robustness and practical deployment, centred around the high-priority research intersection.

Layer	Application	Traditional Performance	ML Performance	Key Enabling Technique
Physical	Localisation accuracy	8.5 m error	0.5–0.8 m error	CNN, DQN active sensing
	Channel estimation MSE	0.043	0.012 (significant reduction)	LSTM temporal modelling
	Modulation classification	75% @ 0 dB SNR	96% @ 0 dB SNR	CNN feature learning
	Adaptive modulation throughput	Baseline	+20–45% (substantial gain)	DQN policy optimisation
MAC	Channel utilisation	8%	18–42% (scenario-dependent)	Q-learning adaptive backoff
	Collision rate	45/hour	12/hour (73% reduction)	Multi-agent RL coordination

TABLE XXVII: Summary of ML Performance Achievements Across IoUT Protocol Stack

Panel of Figure 41 — Fig. 14: Technology roadmap for ML in underwater communications, organising research priorities across four dimensions: data and learning paradigms, computational architectures, collaborative frameworks, and physics integration. Near-term efforts (2025–2027) focus on deployment optimisation using proven techniques; medium-term (2027–2030) enables autonomous adaptation; long-term (2030–2033) achieves intelligent collaboration; and transformative capabilities (2035+) realise the cognitive ocean vision.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Application	Best ML Method	Key Advantages	Constraints	Data Requirements	Accuracy
Physical Layer
Localisation	CNN + k-NN	Sub-metre accuracy, robust to multipath	High memory for fingerprints	1000+ fingerprints	0.8-1.2m
Channel Estimation	LSTM + PINN	Predictive capability, physics-consistent	Computational complexity	100-1000 samples	MSE: 0.012
Modulation Classification	CNN	Robust at low SNR	Requires diverse training	5000+ per class	96% @ 0dB
Adaptive Modulation	DQN	Handles outdated CSI	Large state space	1000+ episodes	20–45% gain
MAC Layer

Appendix

TABLE XXVIII: ML Algorithm Selection Guide for Underwater Applications

Algorithm Class	Memory	FLOPS	Power (W)	Latency (ms)	Recommended Platform
k-NN	O(nd)	O(ndk)	0.01-0.1	10-50	ARM Cortex-M4
Decision Trees	O(nodes)	O(depth)	0.01-0.05	1-10	Any microcontroller
SVM	O(n_sv × d)	O(n_sv × d)	0.05-0.2	5-20	ARM Cortex-M7
Small CNN (¡5 layers)	100KB-1MB	10-100M	0.1-1	10-100	ARM Cortex-A53
Medium CNN (5-20 layers)	1-10MB	100M-1G	1-5	50-500	NVIDIA Jetson Nano
Large CNN (¿20 layers)	10-100MB	1-10G	5-20	100-1000	NVIDIA Jetson Xavier

Appendix

TABLE XXIX: Computational Requirements and Platform Recommendations

Operation	Traditional (J)	ML-Based (J)	Improvement	Battery Life Gain	Key Technique
Acoustic Transmission	10 per packet	0.34 per packet	29×	Weeks → Years	Adaptive power, Q-learning
Channel Estimation	0.5 per estimate	0.08 per estimate	6×	3 → 18 months	CNN prediction
Route Discovery	45 per route	2.1 per route	21×	Days → Months	GNN, caching
Object Detection	8.2 per frame	0.15 per frame	55×	Hours → Days	YOLOv8n, pruning
Network Maintenance	850 per day	12 per day	71×	3 → 214 days	Predictive, federated
Data Compression	2.0 per MB	0.02 per MB	100×	10 → 1000 days	Autoencoder

Appendix

TABLE XXX: Energy Efficiency Comparison: ML vs Traditional Methods

Technology	Complexity	TRL	Time to Deploy	Risk Level	Primary Challenges
Ready for Deployment (TRL 7-9)
k-NN Localisation	Low	8	1-3 months	Low	Training data collection
Q-Learning MAC	Medium	7	3-6 months	Low	Parameter tuning
CNN Channel Est.	Medium	7	3-6 months	Medium	Model size, real-time
Decision Tree	Low	9	¡1 month	Very Low	Limited capability
Pilot Testing (TRL 4-6)

Appendix

TABLE XXXI: Implementation Complexity and Deployment Readiness

Application	Min. Samples	Ideal Samples	Collection Method	Augmentation Strategy	Cost Estimate
Localisation	500	5,000	Grid survey	Noise injection, multipath	$50K-200K
Channel Estimation	100	1,000	Continuous recording	Doppler, time-varying	$20K-100K
Object Detection	1,000	10,000	ROV survey	Colour, turbidity, rotation	$200K-1M
Species Classification	50/class	500/class	Opportunistic + targeted	Pitch shift, time stretch	$100K-500K
Anomaly Detection	1,000 normal	10,000 normal	Long-term monitoring	Synthetic anomalies	$50K-200K
Protocol Learning	100 hours	1,000 hours	Passive recording	Noise, interference	$20K-50K

Appendix

TABLE XXXII: Training Data Requirements and Collection Strategies

Layer 1	Layer 2	Optimisation Method	Performance Gain	Key Insight
Physical	MAC	Joint channel-access learning	35% efficiency	Channel predicts collision probability
Physical	Network	Channel-aware routing	40% reliability	Route around poor channels
MAC	Network	Traffic-aware clustering	45% energy	Cluster based on communication patterns
MAC	Transport	Queue-aware scheduling	60% latency reduction	Prioritise based on transport needs
Network	Transport	Congestion-aware routing	50% throughput	Route around congested nodes
Network	Application	Content-aware routing	30% bandwidth	Different paths for different data types

Appendix

TABLE XXXIII: Cross-Layer Optimisation Opportunities

Environmental Factor	Impact on ML	Adaptation Strategy	ML Technique	Success Rate
Biofouling	Sensor drift, degradation	Progressive calibration	Online learning, EWC	85% maintained
Temperature Variation	Model accuracy drop	Multi-temperature training	Domain adaptation	90% maintained
Pressure (Depth)	Component behaviour change	Depth-stratified models	Ensemble methods	92% maintained
Turbidity	Optical degradation	Robust features	Attention mechanisms	86% maintained
Seasonal Changes	Distribution shift	Continual learning	Progressive networks	88% maintained
Node Mobility	Topology changes	Dynamic retraining	GNN, online RL	91% maintained

Appendix

TABLE XXXIV: Environmental Adaptation Strategies

Investment Area	Initial Cost	Annual OpEx	Benefit/Year	ROI Period	5-Year NPV
Data Collection	$1-5M	$100K	-	-	-$5.5M
Model Development	$0.5-2M	$200K	-	-	-$2.5M
Hardware Upgrade	$5-50K/node	$10K/node	-	-	-$100K/node
Training/Personnel	$200K	$100K	-	-	-$700K
Energy Savings	-	-	$50K/node	Immediate	$200K/node
Maintenance Reduction	-	-	$100K	Year 1	$400K

Appendix

TABLE XXXV: Cost-Benefit Analysis for ML Implementation

Scenario	Network Size	Duration	Budget	Use ML?	Recommended Approach
Short-term monitoring	¡10 nodes	¡1 month	¡$100K	No	Traditional protocols
Coastal surveillance	10-50 nodes	3-12 months	$100K-1M	Partial	ML for critical functions
Long-term monitoring	50-200 nodes	¿1 year	$1-10M	Yes	Full ML stack
Ocean observatory	¿200 nodes	Permanent	¿$10M	Essential	Advanced ML + federation
Research deployment	Any	Variable	Limited	Yes	Transfer learning
Commercial aquaculture	20-100 nodes	Continuous	$500K-5M	Yes	Proven ML solutions

Appendix

TABLE XXXVI: Quick Decision Matrix for ML Adoption

Explore similar work

Sep 14, 2026

Machine Learning for Underwater Optical Wireless Communication Systems: A Comprehensive Survey

Underwater Optical Wireless Communication (UOWC) has emerged as a promising technology for high-speed underwater data transmission, offering significantly higher bandwidth and lower latency compared to acoustic and Radio Frequency (RF) technologies. However, the underwater medium introduces severe impairments that degrade link performance and limit communication range. The growing complexity of these challenges has increased research interest in Machine Learning (ML) and Deep Learning (DL) approaches, which offer powerful tools for channel modeling, signal processing, and system adaptation in ways that conventional analytical methods struggle to achieve. This survey provides a comprehensive review of ML methods applied across the full UOWC system pipeline, covering channel modeling, transmitter design, receiver detection and equalization, link alignment, and emerging applications, including wireless power transfer, semantic communication, object detection, optical sensing, and localization. Finally, we discuss open challenges and future research directions to motivate further work in this rapidly growing field.

Shaymaa Mahmoud, Ardimas Purwita, Mohamed-Slim Alouini

Wireless Communications

Ask PDF

Mar 25, 2026

Energy-Efficient Hierarchical Federated Anomaly Detection for the Internet of Underwater Things via Selective Cooperative Aggregation

Anomaly detection is a core service in the Internet of Underwater Things, yet training accurate distributed models underwater is difficult because acoustic links are low-bandwidth, energy-intensive, and often unable to support direct sensor-to-surface communication. Standard flat federated learning therefore faces two coupled limitations in underwater deployments: expensive long-range transmissions and reduced participation when only a subset of sensors can reach the gateway. This paper proposes an energy-efficient hierarchical federated learning framework for underwater anomaly detection based on three components: feasibility-aware sensor-to-fog association, compressed model-update transmission, and selective cooperative aggregation among fog nodes. The proposed three-tier architecture localises most communication within short-range clusters while activating fog-to-fog exchange only when smaller clusters can benefit from nearby larger neighbours. A physics-grounded underwater acoustic model is used to evaluate detection quality, communication energy, and network participation jointly. In large synthetic deployments, only about 48% of sensors can directly reach the gateway in the 200-sensor case, whereas hierarchical learning preserves full participation through feasible fog paths. Selective cooperation matches the detection accuracy of always-on inter-fog exchange while reducing its energy by 31-33%, and compressed uploads reduce total energy by 71-95% in matched sensitivity tests. Experiments on three real benchmarks further show that low-overhead hierarchical methods remain competitive in detection quality, while flat federated learning defines the minimum-energy operating point. These results provide practical design guidance for underwater deployments operating under severe acoustic communication constraints.

Kenechi Omeke, Michael Mollel, Lei Zhang +2

James Watt School of Engineering, University of Glasgow, U.K.

Federated Learning Hierarchical Federated Learning Internet of Things Federated Learning Aggregation Anomaly Detection

Ask PDF

Mar 27, 2026

ML-Enabled Open RAN: A Comprehensive Survey of Architectures, Challenges, and Opportunities

As wireless communication systems become more advanced, Open Radio Access Networks (O-RAN) stand out as a notable framework that promotes interoperability and cost-effectiveness. An examination of the progression of RAN architectures, as well as O-RAN's underlying principles, reveals the importance of machine learning (ML) in addressing various challenges, including spectrum management, resource allocation, and security. Hence, this survey provides a comprehensive overview of the integration of ML within O-RAN, highlighting its transformative potential in enhancing network performance and efficiency. This survey aims to describe the current status of ML applications in O-RAN while indicating possible directions for future research by analyzing existing literature. The findings aim to assist researchers and stakeholders in formulating optimal service strategies and advancing the understanding of intelligent wireless networks.

Mira Chandra Kirana, Patatchona Keyela, Fatemeh Rostamian +2

The Department of Computer and Software Engineering, Polytechnique Montreal, Montreal, QC, Canada, H3T 1J4

Open Radio Access Network Wireless Communications

Ask PDF