based on the highest probability. With the class probability function of ArcGIS 9.2,
maps were produced showing the probability of a single soil type occurring in a
respective study area.
Scale independent mapping approaches – Classification tree approach: Classification tree-based soil maps using the CART algorithm (Breiman et al. 1984) were
produced for all the research areas and for the whole of north-western Thailand
(Fig. 2.8). The classification trees were generated using SPSS 16.0 (using the
following settings: algorithm CART, parent node: 5, child node: 2 and pruning:
none; growing method: CRT). The input data used were elevation, slope, curvature,
aspect and petrography, plus LANDSAT 7 (bands 1–8) and SPOT 5 (bands 1–4)
data extracted at all the training points.
Scale independent mapping approaches – Random forest approach: ‘Random
forest’ is a classification method which consists of several different, uncorrelated
decision trees, all of which are developed under a certain kind of randomization
during the learning process. For a classification, each tree must make a decision in
this forest and the class with the most votes will decide the final classification
(Breiman 2001; Tin Kam Ho 1995). A random forest-based soil map was produced
for whole of north-west Thailand, with an R-script developed using the package
“randomForest” (Liaw and Wiener 2002) in addition to the package “RODBC”
(Ripley 2012). The used input data comprised a 20 m elevation contour line map
(Royal Thai Survey Department 1976), airborne gamma-ray point data (K, eTh, eU)
provided by DMR, a geological map at the 1:250,000 scale (German Geological
Mission 1979), and point information on the reference soil groups for three different
petrographic areas (Schuler 2008), these being Bor Krai (392 points), Huai Bong
(201 points) and Mae Sa watershed (226 points). SAGA 2.0.8 software was used to
compute a Digital Elevation Model (DEM) from the contour line map and to derive
raster maps of 20 m resolution for elevation, slope, aspect, curvature, profile
curvature, planform curvature, the convergence index, aspect, relative altitude,
the SAGA wetness index and the vertical distance above the channel network.
The same software was also used to produce raster maps of a 50 m resolution for K,
eTh and eU using universal kriging with relative altitude as the covariate. Relative
altitude is a terrain parameter which determines altitude differences within a large
(here 50 km) search radius when compared to each grid cell. This terrain parameter
was computed with a SAGA module by SciLands GmbH. Based on a modified
FAO-parent material classification (FAO 2006), the geological map was converted
into a parent material map and gridded to a 50 m resolution. Finally, all predictor
grids were re-sampled to a 50 m resolution and transferred as application data to a
PostgreSQL database. Data were then extracted at the location of all the training
points and also introduced into the PostgreSQL database. Subsequently, the training
data of all the predictor grids were fitted to the random forest model, and the Giniindex of the predictors was computed. This index represents a measure of the total
decrease in node impurities occurring due to a splitting of the variables, averaged
over all the trees. Afterwards, the model was fitted to the most important predictors
(highest Gini-indices) and applied to the dataset covering the whole area, then the
probability for each reference soil group was computed. The probabilities for each
2 Beyond the Horizons: Challenges and Prospects for Soil Science and Soil. . .
53
maps were produced showing the probability of a single soil type occurring in a
respective study area.
Scale independent mapping approaches – Classification tree approach: Classification tree-based soil maps using the CART algorithm (Breiman et al. 1984) were
produced for all the research areas and for the whole of north-western Thailand
(Fig. 2.8). The classification trees were generated using SPSS 16.0 (using the
following settings: algorithm CART, parent node: 5, child node: 2 and pruning:
none; growing method: CRT). The input data used were elevation, slope, curvature,
aspect and petrography, plus LANDSAT 7 (bands 1–8) and SPOT 5 (bands 1–4)
data extracted at all the training points.
Scale independent mapping approaches – Random forest approach: ‘Random
forest’ is a classification method which consists of several different, uncorrelated
decision trees, all of which are developed under a certain kind of randomization
during the learning process. For a classification, each tree must make a decision in
this forest and the class with the most votes will decide the final classification
(Breiman 2001; Tin Kam Ho 1995). A random forest-based soil map was produced
for whole of north-west Thailand, with an R-script developed using the package
“randomForest” (Liaw and Wiener 2002) in addition to the package “RODBC”
(Ripley 2012). The used input data comprised a 20 m elevation contour line map
(Royal Thai Survey Department 1976), airborne gamma-ray point data (K, eTh, eU)
provided by DMR, a geological map at the 1:250,000 scale (German Geological
Mission 1979), and point information on the reference soil groups for three different
petrographic areas (Schuler 2008), these being Bor Krai (392 points), Huai Bong
(201 points) and Mae Sa watershed (226 points). SAGA 2.0.8 software was used to
compute a Digital Elevation Model (DEM) from the contour line map and to derive
raster maps of 20 m resolution for elevation, slope, aspect, curvature, profile
curvature, planform curvature, the convergence index, aspect, relative altitude,
the SAGA wetness index and the vertical distance above the channel network.
The same software was also used to produce raster maps of a 50 m resolution for K,
eTh and eU using universal kriging with relative altitude as the covariate. Relative
altitude is a terrain parameter which determines altitude differences within a large
(here 50 km) search radius when compared to each grid cell. This terrain parameter
was computed with a SAGA module by SciLands GmbH. Based on a modified
FAO-parent material classification (FAO 2006), the geological map was converted
into a parent material map and gridded to a 50 m resolution. Finally, all predictor
grids were re-sampled to a 50 m resolution and transferred as application data to a
PostgreSQL database. Data were then extracted at the location of all the training
points and also introduced into the PostgreSQL database. Subsequently, the training
data of all the predictor grids were fitted to the random forest model, and the Giniindex of the predictors was computed. This index represents a measure of the total
decrease in node impurities occurring due to a splitting of the variables, averaged
over all the trees. Afterwards, the model was fitted to the most important predictors
(highest Gini-indices) and applied to the dataset covering the whole area, then the
probability for each reference soil group was computed. The probabilities for each
2 Beyond the Horizons: Challenges and Prospects for Soil Science and Soil. . .
53