limestone, mudstone or iron ore as the parent rock materials and a slope inclination
of less than 18
. Alisols occur mostly between 566 and 834 m.a.s.l., with limestone,
mudstone or iron ore as the parent rock materials. Luvisols occur above 834 m.a.s.l.
on latite, while Chernozems and Fluvisols prevail on freshwater limestone or
alluvial deposits, with Chernozems above 734 m.a.s.l. and Fluvisols below this
benchmark. A map for the entire north-west of Thailand could be produced using
this information (Fig. 2.8), but sound validation is still lacking.
Random forest-based soil maps: The random forest model for all the predictor
grids revealed the highest Gini-index values for K (69.8), elevation (61.4), eTh
(58.2) and relative altitude (51.4). Lower values were detected for eU (43.4),
vertical distance to a drainage network (37.7), the wetness index (34.0), curvature
(31.6) and aspect (31.2). Gini-index values for the gamma-ray data stressed the high
level of correlation between radio-element concentration and soil development,
while the low index values for the parent material could be explained by the strong
generalization degree of the geological map. The resulting soil map (Fig. 2.9)
reveals the clear dominance of Acrisols in higher elevation areas and for
consolidated siliceous acid igneous and metamorphic parent materials. In limestone
areas, Alisol soil associations dominate at lower elevations and pass over in Acrisol
dominated areas at higher elevations. Sandstone areas are clearly dominated by
Alisols, with the occasional occurrence of Leptosols. Due to a lack of training data –
as this study focused only on previously rarely studied upland soil associations – it
was not possible to further specify soil associations for the lowland areas. Here,
expert knowledge was provided by Dr. Chaiwong from Maejo University (personal
communication) in northern Thailand, plus the Land Development Department
(1979, 2007). Additional mapping activities will therefore be needed in this area
for validation purposes. A comparison between the classification tree and the
random forest approach for the Bor Krai area revealed that the random forest
approach is more accurate and requires less training points.
2.3.4 Discussion
Six mapping approaches were tested in order to evaluate their strengths and
weaknesses when applied to northern Thailand. At the landform scale, transect
sampling-based soil mapping delivered the best results, but also required the
highest sampling point density to achieve a high level of accuracy. However, this
was anticipated, since the respective reference soil maps were predominantly based
on transect information. The high “type 2 errors” (areas of known RSG X that are
incorrectly classified as anything else – also called ‘errors of omission’ or ‘false
negatives’) with respect to the prediction of minor soil types, showed that only the
transect method could detect the major soil types. The disadvantage of this method
is its restriction to smaller areas if a particular point density is required, and the
difficulty often experienced in selecting the optimum distance between transect
lines.
58
K. Stahr et al.
of less than 18
. Alisols occur mostly between 566 and 834 m.a.s.l., with limestone,
mudstone or iron ore as the parent rock materials. Luvisols occur above 834 m.a.s.l.
on latite, while Chernozems and Fluvisols prevail on freshwater limestone or
alluvial deposits, with Chernozems above 734 m.a.s.l. and Fluvisols below this
benchmark. A map for the entire north-west of Thailand could be produced using
this information (Fig. 2.8), but sound validation is still lacking.
Random forest-based soil maps: The random forest model for all the predictor
grids revealed the highest Gini-index values for K (69.8), elevation (61.4), eTh
(58.2) and relative altitude (51.4). Lower values were detected for eU (43.4),
vertical distance to a drainage network (37.7), the wetness index (34.0), curvature
(31.6) and aspect (31.2). Gini-index values for the gamma-ray data stressed the high
level of correlation between radio-element concentration and soil development,
while the low index values for the parent material could be explained by the strong
generalization degree of the geological map. The resulting soil map (Fig. 2.9)
reveals the clear dominance of Acrisols in higher elevation areas and for
consolidated siliceous acid igneous and metamorphic parent materials. In limestone
areas, Alisol soil associations dominate at lower elevations and pass over in Acrisol
dominated areas at higher elevations. Sandstone areas are clearly dominated by
Alisols, with the occasional occurrence of Leptosols. Due to a lack of training data –
as this study focused only on previously rarely studied upland soil associations – it
was not possible to further specify soil associations for the lowland areas. Here,
expert knowledge was provided by Dr. Chaiwong from Maejo University (personal
communication) in northern Thailand, plus the Land Development Department
(1979, 2007). Additional mapping activities will therefore be needed in this area
for validation purposes. A comparison between the classification tree and the
random forest approach for the Bor Krai area revealed that the random forest
approach is more accurate and requires less training points.
2.3.4 Discussion
Six mapping approaches were tested in order to evaluate their strengths and
weaknesses when applied to northern Thailand. At the landform scale, transect
sampling-based soil mapping delivered the best results, but also required the
highest sampling point density to achieve a high level of accuracy. However, this
was anticipated, since the respective reference soil maps were predominantly based
on transect information. The high “type 2 errors” (areas of known RSG X that are
incorrectly classified as anything else – also called ‘errors of omission’ or ‘false
negatives’) with respect to the prediction of minor soil types, showed that only the
transect method could detect the major soil types. The disadvantage of this method
is its restriction to smaller areas if a particular point density is required, and the
difficulty often experienced in selecting the optimum distance between transect
lines.
58
K. Stahr et al.