EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Understanding catchment influences on flood generating processes - accounting for correlated attributes

Lina Stein1, Martyn Clark2, Francesca Pianosi1, Wouter Knoben2, and Ross Woods1
Lina Stein et al.
  • 1Water and Environmental Engineering, University of Bristol, Bristol, United Kingdom of Great Britain and Northern Ireland (
  • 2Canmore Coldwater Laboratory, University of Saskatchewan, Canmore, Canada

Understanding flood generating mechanisms is critical for model development and evaluation. While several studies analyse how catchment attributes influence flood magnitude and duration, very few studies examine how they influence flood generating processes. Based on prior knowledge about runoff behaviour and flood generation, we assume that flood processes depend not only on climate, but also on catchment characteristics such as topography, vegetation and geology. Specifically, we hypothesize that the influence of catchment attributes on flood processes will vary between different climate types. We tested our hypothesis on the CAMELS dataset, a large sample (671) of catchments in the United States. We classified 61,828 flood events into flood process types using a previously published location-independent classification methodology. Then we quantified the importance of both individual attributes (comparing probability distributions of different flood types) and interacting attributes (using random forests). Accumulated local effects allow interpretability of random forest with correlated attributes. Results show that climate attributes most strongly influence the distribution of flood generating processes within a catchment. However, other catchment attributes can be influential, depending on climate type. Based on the subset of influential catchment attributes, a random forest model can predict flood generating processes with high accuracy for most processes and climates, demonstrating capabilities to predict flood processes in ungauged catchments. Some attributes proved less influential than common hydrologic knowledge would suggest and are not informative in predicting flood process distribution.

How to cite: Stein, L., Clark, M., Pianosi, F., Knoben, W., and Woods, R.: Understanding catchment influences on flood generating processes - accounting for correlated attributes, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-9918,, 2020

Display materials

Display file

Comments on the display material

AC: Author Comment | CC: Community Comment | Report abuse

Display material version 1 – uploaded on 30 Apr 2020
  • CC1: Catchment classification, Matthias Kemter, 05 May 2020

    Dear Lina,

    very interesting study. Do you think that the way you classify the catchments into wet, dry and snow is the best possible solution? The map on slide 4 shows some areas with fluctuating classes for nearby stations (e.g. Florida). Wouldn't a classification by Koppen Geiger regions be more intuitive and homogeneous, while also  allowing for further diversification of classes, if that is desired?

    Matthias Kemter
    Potsdam University

    • AC1: Reply to CC1, Lina Stein, 05 May 2020

      Dear Matthias,

      Thank you for your comment. While I agree that it is a simplistic split into different climate classes, the number of classes was deliberately kept small to ensure large enough sample sizes for the analysis.

      Additionally, the Köppen-Geiger climate classification has been shown to not capture hydrologic variability very well (Knoben et al, 2018).  We split the catchments based on aridity and fraction of precipitation falling as snow, since we assume both attributes have a strong influence on flood process distribution (no snowmelt floods in areas without snowfall). This allows us to evaluate the interaction of these attributes with others.