Do extreme flood events change future flood hazards?

Traditionally flood risk is measured based on trends in flood frequency, driven from streamflow variability, assuming constant channel capacity. Natural disasters, human activities, and changing climate causes non-stationarity in the hydrologic flow regime. In addition, extreme flood events cause erosion of embankments, hillslopes, and terraces, as well as extensive sediment deposition. These spatiotemporal adaptations of the channel properties persist until the fluvial system recovers naturally to its previous state, or it may never recover. During this transition stage, future flood properties (flood depth, frequency, duration, and spatial extent) may change due to the altered channel geometry. In this study, we introduce a machine learning (ML) based framework for studying changes in flood risk following major flood events and demonstrate its predictability based on 3101 USGS stations across the continental US. We train the model with hydrologic (flow, stage), atmospheric (precipitation) and geomorphic (channel width, depth, drainage area, geophysical characteristics) data, and apply it to: 1) identify spatiotemporal interdependencies among atmospheric, geomorphologic, and hydrologic flood drivers; 2) understand the effects of large floods on channel conveyance and consequent effects on future flood hazards; and 3) investigate impacts on future flood hazard for different climate types and geophysical characteristics. This knowledge of the interdependencies of flood drivers will bring new insights on the dynamic changes of flood risk following major flood events.