Abstract

Coastal zones sit at the sharp end of climate stress, dense human settlement and heavy industrial traffic, and the sensing networks we throw at them generate torrents of data that no human team can watch in real time. This paper examines how machine learning driven anomaly detection can be embedded within Internet of Things (IoT) coastal intelligence systems to flag sensor faults, extreme events, pollution incidents and security breaches before they escalate. We review statistical, classical machine learning and deep learning approaches, weigh their behaviour against the awkward realities of marine deployment such as biofouling, intermittent connectivity and severe class imbalance, and set out a layered reference architecture from sensing through to cloud analytics. Comparative evidence drawn from recent benchmarks indicates that reconstruction based deep models, particularly LSTM autoencoders, tend to lead on detection quality, while lightweight ensemble methods remain attractive where edge compute is scarce. We close with the open problems that still block dependable field deployment: label scarcity, concept drift, energy budgets and the trust gap created by opaque models.