ETD

• artificial intelligence
• bandwidth-constrained environments
• edge computing
• edge intelligence
• end-to-end systems
• generative reconstruction
• human-in-the-loop systems
• on-device semantic extraction
• resource-constrained environments
• semantic communication
• semantic representations
• task-oriented communication
• task-oriented metrics
• visual data transmission
• wildlife monitoring

Recent years have witnessed a growing interest in semantic communication, driven by rapid advancements in artificial intelligence and edge computing.
This paradigm aims to improve communication efficiency by prioritizing the transmission of meaning rather than raw bits. Unlike conventional communication systems, which focus on accurate bit-level delivery, semantic communication seeks to convey task-relevant information, potentially reducing bandwidth requirements and increasing robustness through the integration of artificial intelligence and edge computing.
Despite its conceptual appeal, semantic communication is often investigated with a strong emphasis on its potential benefits, while practical limitations, architectural trade-offs, and deployment challenges are only marginally addressed. Moreover, real-world implementations and experimental evaluations in operational and resource constrained environments remain limited.
This thesis investigates the practical viability of semantic communication through the design, implementation, and evaluation of a task-oriented semantic communication system for wildlife monitoring in harsh and bandwidth-constrained scenarios.
The proposed solution is developed within the WatchEDGE follow-up initiative of the RESTART program and targets human-in-the-loop monitoring applications, where visual information must remain interpretable despite severe communication constraints.
An end-to-end system for visual data transmission is presented, combining on-device semantic extraction with generative reconstruction at the receiver. Rather than introducing a new compression codec, the proposed approach explores a semantic redefinition of the transmitted message, where visual content is conveyed through compact semantic representations. As a result, the system achieves a payload reduction exceeding 97% with respect to conventional image transmission pipelines, providing visual data transmission in less than 200 bytes.
Experimental results show that, when evaluated using task-oriented metrics such as detection accuracy and semantic fidelity, the proposed approach can outperform state-of-the-art image compression standards in this specific application context.
Notably, this performance gain was not an a priori objective, but an emergent outcome of the semantic communication design. These results should therefore be interpreted as evidence of the potential of semantic communication, rather than as a general replacement for traditional codecs.
Overall, this work highlights both the opportunities and the current limitations of semantic communication, providing concrete insights into its deployability, the role of human-in-the-loop supervision, and the conditions under which semantic-aware communication can offer tangible benefits over conventional approaches.