Abstract
The stable operation of meteorological stations is crucial for accurate, continuous data that underpin forecasting, disaster prevention, and climate research. Yet existing detectors often miss subtle, localized faults and struggle under real-time, resource-constrained field conditions. We propose Sentinel-Net, an unsupervised, edge-oriented anomaly detector tailored to meteorological station monitoring. Trained only on normal data, Sentinel-Net couples multi-scale feature extraction with a spatial attention module and employs a hybrid MSE-SSIM loss to balance global structure and local detail. For localization, we compute cosine-distance maps, which are robust to global illumination changes and emphasize structural deviations. On real station imagery, Sentinel-Net achieves PSNR 29.72 dB, AUC 0.96, and F1 0.8523 with 9.9 M parameters and 0.07 s/image inference, outperforming classical AE/VAE/AAE (PSNR gains of 23.3%, 28.6%, and 19.2%, respectively) and providing competitive accuracy against recent Mem-AE and Skip-GANomaly while maintaining deployability. These results demonstrate Sentinel-Net's effectiveness and practicality for intelligent meteorological station surveillance.