A multi-source earth observation and artificial intelligence framework for unsupervised water-use anomaly screening in Mediterranean agriculture

Reliable screening of agricultural water-use anomalies over large areas remains challenging because verified event labels and field observations are rarely available. We present AI4WaterGuard, a multi-source Earth Observation and artificial intelligence framework for unsupervised water-use anomaly screening across Cyprus. Using 617,862 records from 1,974 centroids across six hydrological years, six primary detector families were assessed under nested temporal cross-validation with training-consistent scaling, validity-aware scoring, and harmonised temporal support. Among the five centroid-sequence detectors, the convolutional neural network autoencoder showed the strongest response to the strong-reference controlled synthetic proxy, achieving Average Precision of 0.893± 0.130 and Proxy F1 of 0.880± 0.100 . The spatially explicit convolutional long short-term memory variational autoencoder was evaluated at finer land-pixel-sequence granularity and interpreted descriptively. Across five perturbation patterns and three severities, the convolutional neural network autoencoder achieved an equal-scenario mean Average Precision of 0.342, declining to 0.088 for weak perturbations. These synthetic metrics quantify controlled-proxy sensitivity, not field-validated detection performance. Operational post-processing on 1,157 WorldCover-eligible centroids combined threshold sensitivity, 1,000 exact-count randomisations, a primary Top-50 inspection-prioritisation queue with a nested Top-20 high-priority subset, and common-cadence spatiotemporal events. The resulting screening priorities require independent administrative and field evidence to establish physical cause or legal status.