Plain Language Summary
Quantifying past trends in extreme rainfall is important because it can
help us understand future changes caused by global warming. Climate
scientists and hydrologists use specific statistical models to to do so,
but interpreting the results is complicated because extremes are rare
and the structure of the models is not linked to the local meteorology.
We use a new statistical model that allows to better understand the
mechanisms behind the trends we detect. We find that extreme rainfall in
the easter Italian Alps increased over the past decades and we associate
this change to an increased proportion of summer thunderstorms.
1 Introduction
Understanding past and future changes in extreme subdaily precipitation
intensities is of enormous interest because they are responsible for
flash floods, urban floods, landslides and debris flows, and cause
numerous casualties and huge damages every year (Borga et al., 2014;
Cristiano et al., 2017; Paprotny et al., 2018). Physical laws translate
increasing atmospheric temperature into increasing water vapor holding
capacity. Together with changes in the atmospheric dynamics, this is
expected to drive future precipitation changes (Trenberth et al., 2003;
Pendergrass et al., 2020; Fowler et al., 2021b). In general, larger
reponses are expected for precipitation extremes because mean
precipitation, on a global scale, is limited by energy constraints
(Allan and Soden, 2008; Pendergrass & Hartmann, 2014). However,
detecting changes in extreme precipitation is highly affected by the
stochastic uncertainty characterizing the sampling of extremes. This
uncertainty may mask the influence of climate forcing on the processes
which locally control the extremes (Fatichi et al., 2016; Marra et al.,
2019).
Statistically significant changes in the frequency of extreme
precipitation in the past decades were reported, often with stronger
trends in subdaily extremes, as opposed to daily (Guerreiro et al.,
2018; Markonis et al., 2019; Papalexiou & Montanari, 2019). In some
cases, opposing trends between short and long durations emerged, with
complex implications for flood risk (Zheng et al., 2015). Available
observations show different temporal trends for precipitation
intensities associated to different exceedance probabilities (Schär et
al., 2016; Pendergrass, 2018). In general, increasing trends are
reported for rarer events (Myhre et al., 2019), but the specific
differences depend on duration, season, and local conditions, such as
the dominating meteorological features contributing extremes (Blanchet
et al., 2021; Moustakis et al., 2021). Extreme return levels
characterized by different exceedance probabilities are thus changing at
different rates (Myhre et al., 2019; Marra et al., 2021).
Nonstationary extreme value models could aid the detection and
quantification of trends in extreme precipitation of different
exceedance probability (e.g., Min et al. 2009). However, the information
these models can provide is impacted by stochastic uncertainties
(Serinaldi and Kilsby, 2015; Fatichi et al., 2016), and their
flexibility is limited by the assumptions concerning high order
statistical moments. In fact, due to intrinsic limitations in parameter
estimation accuracy, the shape (and sometimes also the scale) parameter
of the extreme value distribution is usually assumed to be constant
(Prosdocimi and Kjeldsen, 2021). Additionally, due to the structure of
these statistical models, a link between the properties of the
underlying process, such as precipitation occurrence frequency and
intensity distribution, and extremes is difficult to establish (e.g.
Marra et al., 2019). As such, the possibility to attribute the observed
changes to specific physical and meteorological processes is hampered.
This background suggests that there is a need to move beyond traditional
trend detection techniques applied to extremes only and develop novel
methodologies. These methods should be able to detect general changes in
extreme precipitation at multiple durations, quantify changes at
different exceedance probabilities, and attribute these to changes in
the underlying physical processes.
Miniussi and Marani (2020) proposed the so-called Metastatical Extreme
Value approach (Marani and Ignaccolo, 2015) as a viable way for
addressing these issues. The idea relies on the concept ofordinary events , that is all the independent realizations of a
process of interest, and proved highly effective in reducing stochasting
uncertainties (Zorzetto et al., 2016; Marra et al., 2018). As opposed to
traditional methods, the distribution describing the ordinary events is
assumed to be known, and the extreme value distribution is derived by
explicitly considering the occurrence frequency of the ordinary events.
Miniussi and Marani (2020)
provided an example application in which extreme return levels where
computed over moving time windows, highlighting temporal changes that
could not be appreciated using traditional methods. The adopted ordinary
events (daily precipitation amounts), however, were not directly
connected with meteorological systems, so that direct relations between
changes in extremes and changes in the underlying storm properties is
still missing.
Here, we combine a novel approach for ordinary-events-based
precipitation frequency analyses across durations (Marra et al., 2020)
with a regional trend detection technique to: (a) detect and quantify
trends in sub-daily annual maxima and extreme return levels by
independently considering the changes in properties and occurrence
frequency of storms, and (b) attribute the observed trends in extremes
to specific changes in the local precipitation regime. We examine the
relevant case of the eastern Italian Alps, where consistent significant
changes in annual maximum precipitation intensities at subdaily and
daily duration were reported (Libertino et al., 2019).
2 Data and methodology
2.1 Study area and data
We focus on Trentino, a 6000 km2-wide mountainous area
in the Eastern Italian Alps (Figure 1 a) which experienced
significant increases in extreme short-duration rain intensities over
the last decades (Libertino et al., 2019). Mean annual precipitation
varies from ~1300 mm yr-1 in the
south-eastern portion of the area to lower amounts (~900
mm yr-1) typical of the “inner alpine province” in
the north (Borga et al., 2005). A dense network of more than one hundred
rain gauges is present. From these, 30 stations (density
~1/200 km-2) with at least 27 complete
years (<10% missing data) of 5-minute resolution data in the
period 1991-2020 are selected (Figure 1 a; see Table S1 in the
Supporting Information).