Outcomes of synthesis - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 371

Figure 12.7. (a) Dependence of cross-validation performance of runoff prediction methods in ungauged basins on data availability (Level 1 and
Level 2 assessments from Chapters 5 - 10). (b) Schematic of spatial hydrological variability (e.g., depth of runoff generation) versus distance
between two points in a region. The schematic illustrates that arid regions tend to exhibit shorter spatial correlation and larger variability than
humid regions, but the average distance of stream gauges tends to be larger.

characteristics. However, one would expect that the
number of stream gauges (or stream gauge density)
impacts differently on predictive performance in different
climates.
In general, arid regions tend to have shorter correlation
lengths and larger variability of runoff-related characteris-
tics than humid regions, as illustrated in the variogram of
Figure 12.7b . Arid regions would therefore need more
gauges to capture the temporal and spatial variability, but
achieving this is unrealistic in many arid parts of the world
where (due to economic reasons) data density is typically
lower than in humid regions. Methods that are able to
exploit the specifics of the region would be needed here.
Use of readily available landscape information, such as
erosional patterns, based on the idea of reading the land-
scape (see Chapter 3 ), may assist in improving runoff
predictions. Examples of the development of novel, low
cost and creative approaches to deal with data scarcity are
presented in the case study chapter ( Chapter 11 ) for various
climates and countries (India, Section 11.2; France,
Section 11.13; Zambia, Section 11.14; Ghana, Section
11.15). As noted in Chapter 3 , a balanced use of remote
sensing data and local data, based on the local experience
of hydrologists, can result

12.2.4 Inter-comparison of methods
Relative performance of different methods
The above analyses have examined general trends of pre-
diction performance with respect to climate/catchment
characteristics and data. This provides a general under-
standing of the controls on performance that goes beyond
individual case studies. However, in the practical context
of estimating runoff in ungauged basins the question is
reversed. Here, one is interested in which method works
best in a given setting. Do some methods naturally lead to
better performance under specific circumstances? The
assessments of Chapters 5 10 have addressed this question

in the comparison of methods diagrams (red diagrams in
Chapters 5

-

10). Table 12.2 provides a summary of these
analyses indicating, for each signature, which two methods
had the highest cross-validation performance for runoff
predictions in ungauged basins. Table 12.3 shows the
methods with lowest performance.
Comparing the best methods from the L1 assessment
with the best methods from the L2 assessment, we find that
(with a few exceptions) the results are almost identical. The
same applies for the methods with lowest performance in
Table 12.3 . This important result indicates that the regions
selected in the L2 assessment are indeed representative of
the wider literature.
Overall, geostatistical methods that include consider-
ation of river network structure seem to work quite well.
These methods capture the way the landscape has evolved
and how water moves through the landscape. However,
geostatistical methods are data-based and work best when
a dense network of stream gauges is available (they work
less well when the network is sparse), which is not the case

-

in a useful strategy for

predictions.
Overall, the comparisons indicate that runoff prediction
performance is strongly related to runoff data availability
and performance is lower in data-poor regions. Concerted
efforts should therefore be made to increase the number
and quality of stream gauges in data-poor regions. Install-
ing a stream gauge is always the best option for determin-
ing runoff.