Spectral Anomaly Benchmark for Long-Term Continuous Observation

SpectrumDet-Anom is a benchmark dataset for evaluating event-level anomaly detection in spectral data. It is constructed from long-term, high-density synchronous acquisition with four types of parameterized, reproducible anomalies injected. The dataset was created by Li, Qin and hosted by Harvard Dataverse, with a last update in April 2026.

Use Cases

• Benchmarking anomaly detection models using the four injected, reproducible event-level anomaly types.
• Evaluating model performance on long-term, continuously observed spectral sequences.
• Testing algorithm robustness against parameterized anomalies within a unified evaluation framework.
• Comparing detection methods for anomalies occurring within regular sampling rhythms.

Strengths

• Designed for long-term continuous observation, supporting extended temporal analysis.
• Provides a unified benchmark with four specific, reproducible types of event-level anomalies.
• Features regular sampling rhythms, enabling consistent time-series evaluation.

Limitations

• The specific sample size, number of features, and temporal range are not provided.
• The dataset's construction is based on injected anomalies, which may not fully represent real-world anomaly distributions.
• Geographic and instrumental sources for the underlying spectral data are unspecified.

Provenance

Source

Harvard Dataverse, authored by Li, Qin.

Collection Method

Constructed from long-term, high-density synchronous spectral acquisition with designed anomaly injection.

Time Range

Described as long-term, but specific dates are unknown.

Freshness

Last updated in April 2026.