Outcomes of synthesis - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 372

Table 12.2. The two methods with the highest cross-validation performance of runoff predictions in ungauged basins
(Level 1 and Level 2 assessments from Chapters 5 10)

-

Ch. 5
Annual

Ch. 6
Seasonal

Ch. 7
FDC

Ch. 8
Low flows

Ch. 9
Floods

Ch. 10
Hydrographs Assessment Catchments

Level 1 all spatial

proximity,
regression

geostat,
regression

short
records,
index

short records,
geostat

geostat,
index

~

Level 2 all index, regional

regression

geostat,
spatial
proximity

geostat,
regression

short records,
process-based

geostat,
index

similarity,
spatial
proximity

Level 2 humid geostat,

spatial
proximity

geostat,
regression

short records,
process-based

geostat,
index

spatial
proximity,
similarity

~

Level 2 arid index, regional

regression

geostat,
process-
based

geostat, regional
regression

geostat,
regression

similarity,
regression

Arid relates to catchments with an aridity index > 1, humid to those with an aridity index < 1.
~ indicates more than two methods with similar performance.

Table 12.3. The two methods with the lowest cross-validation performance of runoff predictions in ungauged basins (Level
1 and Level 2 assessments from Chapters 5 10)

-

Ch. 5
Annual

Ch. 6
Seasonal

Ch. 7
FDC

Ch. 8
Low flows

Ch. 9
Floods

Ch. 10
Hydrographs Assessment Catchments

Level 1 all process-based process-based regression global

regression

regression ~

Level 2 all global

regression

process-based process-

based

global
regression

regression model
average

Level 2 humid ~ regression process-

based

global
regression

regression model
average

Level 2 arid global

regression

regression, spatial
proximity

global
regression

index model

average

Arid relates to catchments with an aridity index 1, humid to those with an aridity index 1.

> <

~ indicates more than two methods with similar performance.

for ungauged (or poorly gauged) basins. Nevertheless,
there is a message here for runoff prediction methods in
ungauged basins. Many current statistical methods ignore
the organisation of catchments by the stream network.
They treat upstream and downstream catchments in the
same way as neighbouring catchments that do not share
the same catchment area. What are needed are methods
that exploit the nested nature of catchments in a consistent
way. A starting point for this may be the top-kriging
approach to predictions in ungauged basins, as discussed
in most of the chapters.
Regression methods generally perform less well than
other methods in this assessment. In fact, regression is
never found to be the best method. Why is this so? Regres-
sion methods hinge on the availability of predictors

(climate and catchment characteristics) that are representa-
tive of the processes reflected by the runoff signatures.
However, it has proved to be generally difficult to
find useful predictors, in particular those that represent
flow paths and storage characteristics of catchments. Very
important parts of the hydrological activity take place
underground (whereas available descriptors tend to be for
the surface landscape), which may help explain why
regression methods do not work as well as other methods.
Clearly, a worthwhile research goal is to find ways to
develop more informative predictors at

the catchment

scale.
In keeping with the above paragraph, global regression
generally has much lower performance than regional
regression. This is, arguably, because it imposes the same