⚡ electrical engineering

Semantic Communication for the Internet of Underwater Things: Architectures, Applications, Challenges, and Future Directions

This paper presents a comprehensive survey on Semantic Communication for the Internet of Underwater Things (IoUT), analyzing its architectures, learning-driven methodologies, and practical applications while identifying critical challenges and future research directions to overcome the unique bandwidth, energy, and latency constraints of underwater networks.

Original authors: Ruhul Amin Khalil, Asiya Jehangir, Hanane Lamaazi, Saddaf Rubab, Nasir Saeed

Published 2026-06-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ruhul Amin Khalil, Asiya Jehangir, Hanane Lamaazi, Saddaf Rubab, Nasir Saeed

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the ocean as a giant, noisy, and very crowded library where everyone is trying to shout messages to each other. This is the Internet of Underwater Things (IoUT). It's a network of underwater cameras, robots, and sensors trying to monitor the environment, inspect pipelines, or track marine life.

The problem? The ocean is a terrible place for shouting.

The "Shout" is slow: Sound travels much slower in water than radio waves do in the air. By the time a message arrives, it might be old news.
The "Room" is small: There is very little "bandwidth" (space to shout). You can't shout a whole movie; you can barely shout a sentence.
The "Noise" is loud: The water is murky, the currents move things around, and the signals get distorted.
The "Battery" is tiny: These underwater devices run on batteries that are hard to change. Shouting constantly drains them fast.

The Old Way vs. The New Way

The Old Way (Raw Data):
Imagine you are an underwater camera. Every second, you take a picture of a fish. In the old system, you try to send the entire high-definition photo to the surface. Even if the fish is just sitting there doing nothing, you send the whole picture. This wastes your battery and clogs the "shouting room" with useless information.

The New Way (Semantic Communication):
This paper proposes a smarter approach called Semantic Communication (SC). Instead of sending the whole photo, the underwater camera uses a little bit of "brain" (AI) to look at the picture first.

The Analogy: Instead of mailing a 500-page book to your boss, you write a one-sentence summary: "The fish is sick."
How it works: The camera extracts the meaning (the "semantics")—like "coral is bleaching," "there is a crack in the pipe," or "a shark is nearby"—and sends just that short message. The receiver (a human or a computer on a ship) gets the point of the message without needing the heavy, raw data.

Why This Matters for the Ocean

The paper argues that this "summary" approach is perfect for the ocean because:

It saves energy: Sending a short text takes much less power than sending a video.
It cuts through the noise: If the message gets a little garbled, the main point ("Pipe broken!") is still clear, whereas a broken video file is useless.
It's faster: You don't have to wait for a huge file to download before you know there's an emergency.

The "Brain" in the Water (The Challenges)

The paper is very careful to say this isn't magic; it's a balancing act.

The Trade-off: To write that one-sentence summary, the underwater robot has to "think" (process the image) first. This uses battery power. The paper asks: Is it worth using battery to think, so we can save even more battery by not shouting? Sometimes, yes. Sometimes, the robot is too small to have a big enough brain to do the thinking.
The "Hallucination" Risk: If the robot uses a fancy AI to guess what it sees, it might make a mistake. It might think a rock is a fish. The paper warns that for safety-critical jobs (like checking a gas pipeline), we can't just trust the AI's guess; we need to be sure.
The "Dictionary" Problem: If Robot A says "Pipe is broken" and Robot B doesn't know what "Pipe" means, they can't talk. The paper suggests we need a standard "dictionary" (like a shared vocabulary) so all underwater devices understand each other.

Real-World Examples Mentioned

The paper lists specific jobs where this "summary" method helps:

Environmental Monitoring: Instead of sending water temperature every minute, the sensor only shouts, "Temperature is rising dangerously!" when something is wrong.
Pipeline Inspection: Instead of sending hours of video of a pipe, the robot sends a map with a red dot saying, "Crack found here, 90% confidence."
Disaster Response: If a tsunami is coming, the system prioritizes the urgent "Tsunami Alert" message over routine data, ensuring the warning gets through first.
Robot Coordination: Two underwater robots can agree on a plan by exchanging simple ideas like "I'm going left, you go right" instead of sharing their entire sensor logs.

The Bottom Line

This paper is a roadmap. It says, "We have a great idea to make underwater communication smarter by sending meaning instead of data." However, it also admits we aren't there yet. We need to:

Build better, smaller "brains" for underwater robots that don't eat too much battery.
Create standard rules so all devices speak the same language.
Test these ideas in real oceans, not just on computers, to make sure they work when the water is muddy and the signals are weak.

In short, the paper suggests we stop trying to shout the whole ocean to the surface and start sending the most important headlines instead.

Based on the provided text, here is a detailed technical summary of the paper "Semantic Communication for the Internet of Underwater Things: Architectures, Applications, Challenges, and Future Directions."

Problem Statement

The Internet of Underwater Things (IoUT) faces severe communication constraints that render traditional raw-data transmission inefficient. These constraints include limited bandwidth (particularly in acoustic links), long propagation delays, time-varying channels, intermittent connectivity, and strict energy budgets due to the difficulty of underwater maintenance. While IoUT supports critical applications like environmental monitoring and subsea inspection, the volume of data generated by sensors, cameras, and Autonomous Underwater Vehicles (AUVs) often exceeds the capacity of available links. Transmitting raw data (e.g., full-resolution images or continuous sensor streams) consumes excessive bandwidth and energy, often delivering information that is irrelevant to the specific mission task.

Methodology and Framework

This paper presents a critical and feasibility-aware survey of Semantic Communication (SC) as a paradigm shift for IoUT. Rather than focusing on the accurate transmission of bits, SC aims to transmit task-relevant meaning. The authors structure their analysis around several core components:

Semantic Representation: The paper defines the semantic encoder as a function $s = f_\theta(x, c, g)$ , where raw observation $x$ is transformed into a compact representation $s$ based on context $c$ and task goal $g$ . Representations can range from symbolic labels and knowledge-graph triples to feature embeddings and anomaly alerts.
Layered Architecture: A four-layer SC architecture is proposed:
- Sensing/Device Layer: Performs local preprocessing and semantic extraction (e.g., event detection) before transmission.
- Network/Transport Layer: Forwards semantic packets with metadata (urgency, confidence, provenance) using semantic-aware routing and link selection.
- Processing Layer: Handles semantic fusion, reasoning, and distributed learning (e.g., Federated Learning) at edge gateways or surface stations.
- Application Layer: Converts semantic information into actionable decisions, prioritizing task utility over bit-level fidelity.
Learning-Driven Approaches: The survey categorizes SC techniques into:
- Data-Driven: Deep Learning (DL) models like CNNs, Transformers, and Deep Joint Source-Channel Coding (Deep JSCC) that learn representations from data.
- Knowledge-Driven: Use of ontologies and Knowledge Graphs (KGs) to enforce domain rules and consistency.
- Hybrid: Combining neural perception with symbolic reasoning to handle uncertainty and safety constraints.
- Generative Models: Using Generative AI (e.g., diffusion models) at the receiver to reconstruct approximate visualizations from compact descriptors, though the paper notes the risk of "hallucinations" in safety-critical tasks.
- Federated Learning (FL): Enabling collaborative model training across distributed IoUT nodes without sharing raw data, addressing privacy and bandwidth constraints.
Semantic Channel Modeling: The paper extends traditional channel models to include semantic distortion, model mismatch, and context errors. It introduces metrics like Semantic Outage Probability (the probability that task-level distortion exceeds a threshold) and Semantic Reliability, distinguishing them from traditional Bit Error Rate (BER).

Key Contributions

The paper makes the following specific contributions:

Focused Survey: It provides the first focused synthesis of SC specifically for IoUT, analyzing how underwater constraints (acoustic/optical heterogeneity, energy limits) impact the design of meaning-driven systems.
Structured Taxonomy: It develops a taxonomy of SC-enabled IoUT systems based on representation type (symbolic, feature-level, generative), communication modality, and learning paradigm.
Feasibility Analysis: It critically examines the compute–communication trade-off, demonstrating that semantic extraction is only beneficial if the energy saved in transmission exceeds the energy cost of local inference and model updates.
Application Analysis: It analyzes representative IoUT applications (environmental monitoring, archaeology, pipeline inspection, disaster response, AUV coordination) from an SC perspective, identifying specific semantic units and risks for each.
Future Directions: It identifies key research gaps, including the need for standardized semantic protocols, reproducible testbeds, lightweight edge AI, semantic security, and digital twin integration.

Results and Illustrative Findings

The paper does not present new experimental field data but relies on a structured review and illustrative conceptual simulations to demonstrate trends:

Efficiency Trade-offs: Illustrative simulations (Figs. 5 and 7) suggest that while raw bit-oriented transmission requires high Signal-to-Noise Ratio (SNR) for utility, SC descriptor transmission can achieve higher energy-normalized semantic efficiency at lower SNRs, provided the inference cost is manageable.
Reliability vs. Utility: The analysis shows that packet-level reliability (e.g., Packet Delivery Ratio) does not guarantee semantic reliability. A packet may be delivered error-free but fail to convey the correct task meaning due to model mismatch or context errors.
Model Deployment: The paper highlights that large models (Transformers, Vision-Language Models) are generally unsuitable for low-power underwater sensor nodes due to memory and energy constraints, advocating for a hierarchical deployment where lightweight models run on sensors and complex reasoning occurs at surface/cloud nodes.
Generative Risks: While generative models can reduce payload sizes significantly, the paper emphasizes that their outputs are probabilistic and may contain incorrect details, making them unsuitable as sole evidence for safety-critical decisions without confidence calibration and fallback mechanisms.

Significance and Claims

The paper claims that SC is a promising but not universal solution for IoUT. Its significance lies in shifting the design philosophy from "delivering all data" to "delivering useful meaning."

Resource Optimization: By transmitting only task-relevant information, SC can significantly alleviate bandwidth and energy constraints in underwater networks.
Task-Oriented Design: It argues that communication success in IoUT should be measured by the utility of the information for a downstream decision (e.g., detecting a pipeline leak) rather than traditional metrics like throughput or BER.
Cautionary Note: The authors explicitly state that SC introduces new challenges, including semantic security risks (leakage of meaning even if raw data is hidden), the need for standardized ontologies, and the risk of generative hallucinations. They conclude that rigorous SC-enabled IoUT design must jointly consider bandwidth savings, edge-computation costs, model size, semantic fidelity, and security.

The paper positions itself as a foundational guide for researchers moving from theoretical semantic concepts to practical, trustworthy, and resource-aware underwater communication systems.