Process oriented insights from interpretable machine learning - what
influences flood generating processes?
Abstract
Hydroclimatic flood generating processes, such as excess rain, short
rain, long rain, snowmelt and rain-on-snow, underpin our understanding
of flood behaviour. Knowledge about flood generating processes helps to
improve modelling decisions, flood frequency analysis, estimation of
climate change impact on floods, etc. Yet, not much is known about how
climate and catchment attributes influence the distribution of flood
generating processes. With this study we aim to offer a comprehensive
and structured approach to close this knowledge gap. We employ a large
sample approach (671 catchment in the conterminous United States) and
test attribute influence on flood processes with two complementary
approaches: firstly, a data-based approach which compares attribute
probability distributions of different flood processes, and secondly, a
random forest model in combination with an interpretable machine
learning approach (accumulated local effects). This machine learning
technique is new to hydrology, and it overcomes a significant obstacle
in many statistical methods, the confounding effect of correlated
catchment attributes. As expected, we find climate attributes (fraction
of snow, aridity, precipitation seasonality and mean precipitation) to
be most influential on flood process distribution. However, attribute
influence varies both with process and climate type. We also find that
flood processes can be predicted for ungauged catchments with relatively
high accuracy (R2 between 0.45 and 0.9). The implication of these
findings is that flood processes should be taken into account for future
climate change impact studies, as impact will vary between processes.