3D Convolutional Autoencoder with Spectral–Spatial Attention (3DCAE-SSA) for Anomaly Detection in Hyperspectral Imagery

This study evaluates how combining weather models with satellite navigation and meteorological data improves forecasts for extreme events, using the historic July 2021 European floods as a test case. We show that while assimilation significantly boosts our ability to detect rainfall and corrects near-surface temperature errors by up to 97 %, it does not necessarily improve the prediction of rainfall intensity. This has key implications for designing better regional flood warning systems.

Artificial intelligence weather models are now competing with traditional physics-based models for flood forecasting. We tested three such models and one physics-based model on three Luxembourg flood events. The artificial intelligence models detected rain and predicted temperatures better, but only the physics-based model reproduced the deadly 2021 European floods. Their skill held for events outside their training data. Both types together offer the most reliable approach for flood warning.