Prediction of floods in ungauged basins - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 225

- There is a tendency for the performance to increase

with catchment elevation.

- There is a clear increase in the performance of all methods

with catchment size. In larger catchments runoff data
from subcatchments or close downstream neighbours
are often available and aggregation effects may make
floods more predictable than in smaller catchments.

- In humid conditions index methods and geostatistical

methods perform slightly better than regression methods.
However, in arid conditions the index methods are sig-
nificantly biased and significantly overestimate the 100-
year floods in the catchments analysed.

9.6 Summary of key points

The flood frequency curve is a particular signature of
runoff variability that describes the (inter-annual) distri-
bution of the annual maximum runoff. The flood fre-
quency curve arises through the interaction of event
rainfall variability with several catchment processes
(e.g., runoff generation, runoff routing, evaporation as
controlled by antecedent soil moisture).

The flood frequency curve reflects the distribution of
rainfall in time (duration, intensity, frequency) and space
(patchiness, orographic effects, storm movement), the
distribution of water flow paths (surface, subsurface,
channel), the seasonality of climate and the resulting
soil moisture variations, and the interplay of all of these.
These are then the process elements that impact the
flood frequency curve.

However, floods shape landscapes through soil erosion
and deposition, the generation and maintenance of
river networks, and associated soil and vegetation pat-
terns. Just as in the case of geomorphology (Haff, 1996 ),
all of these can be deemed emergent patterns, and can
therefore serve as predictors of flood frequency.
Examples of co-evolutionary or emergent variables as
predictors include mean annual precipitation and drain-
age density (because they explain, over long time scales,
both the event characteristics, antecedent soil moisture
and drainage patterns), and the hypsometric curve,
which characterises the distribution of elevation within
a catchment.

Figure 9.30. Absolute normalised error (ANE) of predicting the
100-year flood in ungauged basins for different regionalisation
methods, stratified by aridity. Lines connect median absolute
normalised errors for the same study. Boxes are 40% - 60% quantiles,
whiskers are 20% - 80% quantiles. After Salinas et al.( 2013 ).

1.0, indicating that typically the methods predict twice the
floods actually observed. If a homogeneous region con-
tains both arid catchments with relatively lower floods
and wetter catchments with higher floods, the homogen-
eity assumption will tend to lead to an overestimation in
those catchments with the lower floods. The other two
methods, however, remain unbiased for the most arid
catchments.

Over the past decades there has been an expansion in the
use of more temporal, spatial and causal information in
flood frequency estimation. With the passing of time it is
now possible to identify long-term trends in the data.
The expansion of the spatial scope of flood frequency
estimation enables one to identify spatial trends.
Increased availability of data and improved process
understanding help to better interpret and learn from
these spatial and temporal trends. Also, there is an
increasing tendency, in data-rich regions, to utilise the

Main findings of Level 2 assessment
- The flood prediction performance of all methods for

ungauged basins decreases with increasing aridity
and air temperature.