Interpretable Machine Learning Reveals Flood Generation Processes of Catchments in China
Abstract
Current definitions of flood generation processes explain why flood events happen, while it do not quantify how event indicators cause different magnitudes of flood peaks. In this study, we use interpretable machine learning to quantify the relationships between flood peaks and event indicators, which reveals the flood generation processes of catchments. First, we build a data-driven model to predict flood event peaks using a comprehensive set of event indicators, i.e., the magnitudes, spatial and temporal organizations of various hydrological drivers (e.g., rainfall, snowmelt, soil moisture, and baseflow). Then, interpretable machine learning techniques are used to summarize the response patterns of flood peaks to different event indicators in three types of events, i.e., rainfall, snowmelt, and mixture events. The new method is applied in around 400 catchments with at least 10 years of daily streamflow data in China. Results suggest that hydrological drivers tend to be spatially and temporally uniform. Therefore, small and large floods are mainly distinguished by the magnitudes of rainfall and snowmelt. The response patterns have no significant spatial clustering , which means even catchments that are closed in space may have different sensitivities of flood peaks to event indicators. This study shows how flood peaks change with event indicators to define flood generation processes. In further studies, the new perspective of flood generation processes will benefit flood change attribution and regional flood frequency analysis.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H41A..04Y