The grid-based randomized sampling approach showed similar inherent
difficulties, with the main problems being the choice of an appropriate grid size
and the minimum number of randomized points per grid cell requiring ex-ante
knowledge of local soil variability. The concept of this approach assures an
objective mapping of the main soil types and properties, and the irregular distances
set for the sampling points might also be an advantage for geo-statistical approaches
based on kriging. Again, a high number of “type 2 errors” for the minor soil types
occurred here, suggesting the same weakness as seen in the catena approach.
A technical problem experienced here was the location of the preselected random
points in often steep and inaccessible terrain.
The local knowledge-based soil mapping approach proved to be very rapid and
cost efficient, though the boundaries between mapping units often did not correspond with those on the WRB soil map, because local people applied different
criteria according to their needs and concerns. Classification may even vary within
villages (Schuler et al. 2006) depending on the informants and approach used,
making validation as vital as with other mapping approaches. Nevertheless, local
soil maps provide a good overview about soil type and soil property diversity, and
are suitable as the basis for reconnaissance surveys, land-use planning and to feed
expert systems, plus facilitate optimal grid size selection, sample point density and
the selection of transect distance for the other mapping approaches.
The maximum likelihood-based maps used showed a high level of
correspondences with the reference soil maps and the independent sampling points,
with at least a 70 % level of correct matches found. The high number of errors
among less common soil types indicated that this method is at the moment only
applicable for major soil types on a smaller scale (¼ greater areas), but it may still
be possible to improve this approach by introducing other information on soil
forming factors, such as high resolution climatic data. One advantage of the
maximum likelihood-based soil mapping approach is that up-scaling from a small
calibration area to a surrounding larger target area is easy, provided that petrography, topography and land-use patterns in both areas are similar. Under such
conditions, the target area can exceed – as our study demonstrated for the Mae Sa
watershed – the calibration area by more than 10 times. As for the grid-based
randomized mapping approach, ex-ante knowledge of the expected soil variability
is an advantage when wishing to optimize sampling point numbers.
Carrying out ex-ante principal component analysis (PCA) on the available data
layers in order to provide maximum likelihood classification greatly improved the
outcome, and according to the weighting factors provided by PCA, elevation
seemed to have a minor predictive power, but omitting this variable led to very
poor results. In contrast, the elimination of the more highly weighted SPOT 5 band
4 caused only a slight degradation in the results. Therefore, PCA weighting should
not be overestimated, but additional sensitivity analysis conducted when using it.
PCA also revealed the significant importance of aspect, which might have been due
to the prevalence of north–south trending mountain ranges in the area and also that
aspect is related to the presence of microclimates, which can be expressed through
vegetation moisture differences and are quite pronounced in the Mae Sa watershed
2 Beyond the Horizons: Challenges and Prospects for Soil Science and Soil. . .
59
difficulties, with the main problems being the choice of an appropriate grid size
and the minimum number of randomized points per grid cell requiring ex-ante
knowledge of local soil variability. The concept of this approach assures an
objective mapping of the main soil types and properties, and the irregular distances
set for the sampling points might also be an advantage for geo-statistical approaches
based on kriging. Again, a high number of “type 2 errors” for the minor soil types
occurred here, suggesting the same weakness as seen in the catena approach.
A technical problem experienced here was the location of the preselected random
points in often steep and inaccessible terrain.
The local knowledge-based soil mapping approach proved to be very rapid and
cost efficient, though the boundaries between mapping units often did not correspond with those on the WRB soil map, because local people applied different
criteria according to their needs and concerns. Classification may even vary within
villages (Schuler et al. 2006) depending on the informants and approach used,
making validation as vital as with other mapping approaches. Nevertheless, local
soil maps provide a good overview about soil type and soil property diversity, and
are suitable as the basis for reconnaissance surveys, land-use planning and to feed
expert systems, plus facilitate optimal grid size selection, sample point density and
the selection of transect distance for the other mapping approaches.
The maximum likelihood-based maps used showed a high level of
correspondences with the reference soil maps and the independent sampling points,
with at least a 70 % level of correct matches found. The high number of errors
among less common soil types indicated that this method is at the moment only
applicable for major soil types on a smaller scale (¼ greater areas), but it may still
be possible to improve this approach by introducing other information on soil
forming factors, such as high resolution climatic data. One advantage of the
maximum likelihood-based soil mapping approach is that up-scaling from a small
calibration area to a surrounding larger target area is easy, provided that petrography, topography and land-use patterns in both areas are similar. Under such
conditions, the target area can exceed – as our study demonstrated for the Mae Sa
watershed – the calibration area by more than 10 times. As for the grid-based
randomized mapping approach, ex-ante knowledge of the expected soil variability
is an advantage when wishing to optimize sampling point numbers.
Carrying out ex-ante principal component analysis (PCA) on the available data
layers in order to provide maximum likelihood classification greatly improved the
outcome, and according to the weighting factors provided by PCA, elevation
seemed to have a minor predictive power, but omitting this variable led to very
poor results. In contrast, the elimination of the more highly weighted SPOT 5 band
4 caused only a slight degradation in the results. Therefore, PCA weighting should
not be overestimated, but additional sensitivity analysis conducted when using it.
PCA also revealed the significant importance of aspect, which might have been due
to the prevalence of north–south trending mountain ranges in the area and also that
aspect is related to the presence of microclimates, which can be expressed through
vegetation moisture differences and are quite pronounced in the Mae Sa watershed
2 Beyond the Horizons: Challenges and Prospects for Soil Science and Soil. . .
59