NightWatch: Foundation Model based method for anomaly detection in NightTime Lights

Nighttime light (NTL) satellite imagery provides a valuable signal for monitoring human activity, and has been used to track events such as urbanization, economic development, disasters, and conflicts. The latter are often visible through temporal anomalies; however, NTL time series exhibit strong variability driven by moonlight, cloud cover, atmospheric conditions, and sensor geometry, which makes it difficult to distinguish genuine anomalies from natural fluctuations. Most existing anomaly and change point detection methods operate offline, requiring the full time series in advance, which limits their applicability to online, real-world monitoring settings. Morevoer, the few methods designed specifically for NTL require training a separate model for each region, which limits their scalability and generalization ability to new regions and event types. In this paper, we introduce NightWatch, an online anomaly detection method that leverages a pretrained time series foundation model to flag events in real time without region-specific retraining. We further introduce a new dataset of NTL time series with annotated events spanning multiple event types and regions, and use it to compare NightWatch against classical online detection methods and different foundation model backbones. Our results show that NightWatch detects events with low delay and generalizes across regions and event types without retraining, making it well suited for real-world NTL monitoring.