variable. The difference between the two was then averaged over all trees, and
normalized by the standard deviation of the differences (Liaw 2012). The importance lies in the increase in the value (Table 2.4).
Generally, the most important predictors for the whole of north-west Thailand
were K, eTh and elevation (dem), underlining the potential of using gammaspectrometry for small-scale soil mapping. This is due to the fact that K, eTh and
eU can be considered as proxies for the soil parent material (as one soil forming
factor), as well as for specific soil genetic processes like clay illuviation, at quite a
high level of resolution (see next section), while elevation is a proxy for the local
climate (another important soil forming factor). The soundness of the weighting
used in the random forest approach can be demonstrated by two examples:
i. Regosols showed the highest variability of radio-elements, due to the fact that
Regosols are young soils which are still dominantly characterized by their parent
material, and ii. Fluvisols were mainly determined by elevation, which is obvious
since they appear in the lowest elevation landscape positions.
2.3.5 Conclusions Regarding Soil Mapping Procedures
In this study, the suitability of the different mapping approaches depended mainly
on the scale of the intended application. From the field to the sub-watershed scales,
the transect-based mapping approach delivered the highest resolution and most
accurate results; however, this approach is time-consuming, labor intensive and not
so suitable when wishing to map larger areas. The alternatively applied randomized
grid-based mapping approach produced quite satisfying results, but cannot be
applied in difficult terrain. The cheapest and quickest approach to use for soil
mapping at the village scale is to elicit local soil knowledge, as this study revealed
that some farmers had a good level of knowledge about the local soils and their
properties. Local soil maps offer high levels of potential in terms of land-use
planning, but local soil classifications should be restricted to village areas only,
and cannot easily be transferred into international soil classification systems.
Nevertheless, local soil knowledge can be reasonably included in composite
mapping approaches, in order to obtain a rapid overview of soil diversity and derive
the necessary sampling density. At scales exceeding the sub-watershed level, the
application of the above mentioned mapping approaches is not suitable due to time
and workload constraints. The maximum likelihood approach ranges in accuracy
for the same level as the classification tree approach, but offers the opportunity to
up-scale. However, it is still questionable as to whether it can be applied to areas
containing high petrographic variability, as it requires many training points to be set
up and is not robust due to noise in the data. In the classification tree-based map
used here, mapping unit boundaries still correlated well with petrographic units;
however, it remains uncertain as to whether mapping rules established via restricted
training areas can be transferred across the whole region. For example, in most
cases elevation is a substitute for the rainfall/evapotranspiration ratio, and regional
62
K. Stahr et al.
normalized by the standard deviation of the differences (Liaw 2012). The importance lies in the increase in the value (Table 2.4).
Generally, the most important predictors for the whole of north-west Thailand
were K, eTh and elevation (dem), underlining the potential of using gammaspectrometry for small-scale soil mapping. This is due to the fact that K, eTh and
eU can be considered as proxies for the soil parent material (as one soil forming
factor), as well as for specific soil genetic processes like clay illuviation, at quite a
high level of resolution (see next section), while elevation is a proxy for the local
climate (another important soil forming factor). The soundness of the weighting
used in the random forest approach can be demonstrated by two examples:
i. Regosols showed the highest variability of radio-elements, due to the fact that
Regosols are young soils which are still dominantly characterized by their parent
material, and ii. Fluvisols were mainly determined by elevation, which is obvious
since they appear in the lowest elevation landscape positions.
2.3.5 Conclusions Regarding Soil Mapping Procedures
In this study, the suitability of the different mapping approaches depended mainly
on the scale of the intended application. From the field to the sub-watershed scales,
the transect-based mapping approach delivered the highest resolution and most
accurate results; however, this approach is time-consuming, labor intensive and not
so suitable when wishing to map larger areas. The alternatively applied randomized
grid-based mapping approach produced quite satisfying results, but cannot be
applied in difficult terrain. The cheapest and quickest approach to use for soil
mapping at the village scale is to elicit local soil knowledge, as this study revealed
that some farmers had a good level of knowledge about the local soils and their
properties. Local soil maps offer high levels of potential in terms of land-use
planning, but local soil classifications should be restricted to village areas only,
and cannot easily be transferred into international soil classification systems.
Nevertheless, local soil knowledge can be reasonably included in composite
mapping approaches, in order to obtain a rapid overview of soil diversity and derive
the necessary sampling density. At scales exceeding the sub-watershed level, the
application of the above mentioned mapping approaches is not suitable due to time
and workload constraints. The maximum likelihood approach ranges in accuracy
for the same level as the classification tree approach, but offers the opportunity to
up-scale. However, it is still questionable as to whether it can be applied to areas
containing high petrographic variability, as it requires many training points to be set
up and is not robust due to noise in the data. In the classification tree-based map
used here, mapping unit boundaries still correlated well with petrographic units;
however, it remains uncertain as to whether mapping rules established via restricted
training areas can be transferred across the whole region. For example, in most
cases elevation is a substitute for the rainfall/evapotranspiration ratio, and regional
62
K. Stahr et al.