Public Emergency Alerts as Weak Labels for Weather Station Anomaly Detection: Evaluating an LSTM Autoencoder in Costa Rica

Low-cost weather station networks rarely have labelled anomalies, so detectors are validated against proxy labels or not at all. We present CLIMA-μEMA (Climate Localised Incident Monitoring with Autoencoders), a pipeline that turns the public alerts of Costa Rica’s National Emergency Commission into weak labels for an LSTM-autoencoder anomaly detector on the μEMA network, and test whether such labels can validate it. Under the standard point-adjusted protocol the detector reaches F1 of 0.94–0.98 at five stations, but a wet-season calendar rule that uses no sensor data scores higher at all five, with no false positives. Test sets are 78–86% positive, and 17–22 alert episodes per station merge into one or two credited blocks. With season and time period held fixed, ranking is near chance (median AUC 0.49) under raw or curated labels. Region-level alerts thus encode season and test-set construction rather than events; such evaluations should report prevalence, block counts and calendar baselines.