Estimating the frequency and magnitude of natural hazards largely hinges on stationary models, which do not account for changes in the climatological, hydrological, and geophysical baseline conditions. Using five diverse case studies encompassing various natural hazard types, we present advanced statistical and machine learning methods to analyze and model transient states from long-term inventory data. A novel storminess metric reveals increasing European winter windstorm severity from 1950 to 2010. Non-stationary extreme value models quantify trends, seasonal shifts, and regional differences in extreme precipitation for Germany between 1941 and 2021. Utilizing quantile sampling and empirical mode decomposition on 148 years of daily weather and discharge data in the European Alps, we assess the impacts of changing snow cover, precipitation, and anthropogenic river network modifications on river runoff. Moreover, a probabilistic framework estimates return periods of glacier lake outburst floods in the Himalayas, demonstrating large differences in 100-year flood levels. Utilizing a Bayesian change point algorithm, we track the onset of increased seismicity in the southern central United States and find correlation with wastewater injections into deep wells. In conclusion, data science reveals transient states for very different natural hazard types, characterized by diverse forms of change, ranging from gradual trends to sudden change points and from altered seasonality to overall intensity variations. In synergy with the physical understanding of Earth science, we gain important new insights into the dynamics of the studied hazards and their possible mechanisms.
