Abstract
Cultural heritage (CH) sites are continuously exposed to pressures that can lead to deterioration. Continuous monitoring and anomaly detection (AD) enable early damage detection for preventive conservation. We propose a late-fusion AD framework combining NST-Net, a nested spatiotemporal neural network, with a semantic module encoding expert conservation rules, and evaluate it on five real-world CH monitoring datasets. NST-Net outperforms six state-of-the-art baselines on most sites, achieving, on average, 42% higher Average Precision than the best baseline per site under a synthetic evaluation protocol emphasizing sharp, short-duration anomalies. Dataset length serves as an important performance factor: the combined architecture and preprocessing pipeline particularly benefit longer campaigns, whereas they add little value on shorter ones. NST-Net also requires approximately 13 times less peak memory than baseline detectors, at a higher but still millisecond-scale latency. Further analysis shows that performance depends strongly on anomaly density, type, and severity. The semantic module passes label-free sanity checks and complements NST-Net effectively (complementarity index > 0.87), identifying conservation-relevant events missed by the deep model. These findings support late-fusion statistical–semantic frameworks for AD in CH monitoring.