unnecessary predictors tend to be ignored. The more complex random forest
approach is harder to interpret when compared to classification tree mapping, but
requires less training points to perform a soil type prediction of a higher accuracy.
This was proven in the study by use of only a few Fluvisol points in the Bor Krai
area to calibrate both random forest and classification trees. While the random
forest approach perfectly predicted Fluvisols for the valley bottom in the Nam Lan
valley south of Bor Krai, the classification tree predicted Fluvisols for the valley
bottom, but also quite unrealistically for the adjacent escarpment – to elevations of
up to 100 m above the river (compare Figs. 2.8 and 2.9). The random forest
approach was also applied across the whole of north-west Thailand, and the
resulting soil map corresponds to a very high degree with the reference maps for
the training areas. Within the frame of the random forest approach, the importance
of potential predictors for the RSGs was determined using the mean decrease in an
accuracy measure, one computed from the permuting of out of bag (OOB) data
(which corresponded to about one-third of the cases left out of the sample when the
training set for the current tree in Random Forest was drawn – by sampling with the
replacement). This data were used to get a running unbiased estimate of
the classification error, as trees were added to the forest, plus were used to estimate
the variables’ importance. For each tree, the prediction error on the OOB portion
of the data was recorded (error rate for classification; mean square errors (MSE) for
regression). After that, the same was carried out after permuting each predictor
Table 2.3 Satellite data properties used in the study
Satellite/band
Resolution
[m]
Range [m]
Detection/
application
(globally)
Detection/application
(NW-Thailand)
LANDSAT
7 ETM+
5 30
Near Infrared:
1.55–1.75
Vegetation
moisture, soil
moisture,
differentiation of
snow from
clouds
Bor Krai: discrimination of
Alisols and Acrisols
Huai Bong:
discrimination of
Alisols, Cambisols,
Regosols and Leptosols
7 30
Mid Infrared:
2.08–2.35
Minerals and rock
types; vegetation
moisture
Bor Krai: discrimination of
Alisols and Acrisols
Huai Bong:
discrimination of
Alisols, Cambisols,
Regosols and Leptosols
SPOT 5
2 10
Visible (red):
0.61–0.68
Roads, bare soil;
discrimination of
vegetated/nonvegetated areas
All areas: discrimination of
Alisols, Acrisols,
Cambisols and
Technosols
3 10
Near Infrared:
0.78–0.89
Vegetation biomass,
water-vegetation
discrimination
Mae Sa Mai: discrimination
of Acrisols and
Cambisols
4 2.5
Panchromatic:
0.49–0.69
Provides higher
resolution
All areas: discrimination of
Alisols, Acrisols and
Cambisols
2 Beyond the Horizons: Challenges and Prospects for Soil Science and Soil. . .
61
approach is harder to interpret when compared to classification tree mapping, but
requires less training points to perform a soil type prediction of a higher accuracy.
This was proven in the study by use of only a few Fluvisol points in the Bor Krai
area to calibrate both random forest and classification trees. While the random
forest approach perfectly predicted Fluvisols for the valley bottom in the Nam Lan
valley south of Bor Krai, the classification tree predicted Fluvisols for the valley
bottom, but also quite unrealistically for the adjacent escarpment – to elevations of
up to 100 m above the river (compare Figs. 2.8 and 2.9). The random forest
approach was also applied across the whole of north-west Thailand, and the
resulting soil map corresponds to a very high degree with the reference maps for
the training areas. Within the frame of the random forest approach, the importance
of potential predictors for the RSGs was determined using the mean decrease in an
accuracy measure, one computed from the permuting of out of bag (OOB) data
(which corresponded to about one-third of the cases left out of the sample when the
training set for the current tree in Random Forest was drawn – by sampling with the
replacement). This data were used to get a running unbiased estimate of
the classification error, as trees were added to the forest, plus were used to estimate
the variables’ importance. For each tree, the prediction error on the OOB portion
of the data was recorded (error rate for classification; mean square errors (MSE) for
regression). After that, the same was carried out after permuting each predictor
Table 2.3 Satellite data properties used in the study
Satellite/band
Resolution
[m]
Range [m]
Detection/
application
(globally)
Detection/application
(NW-Thailand)
LANDSAT
7 ETM+
5 30
Near Infrared:
1.55–1.75
Vegetation
moisture, soil
moisture,
differentiation of
snow from
clouds
Bor Krai: discrimination of
Alisols and Acrisols
Huai Bong:
discrimination of
Alisols, Cambisols,
Regosols and Leptosols
7 30
Mid Infrared:
2.08–2.35
Minerals and rock
types; vegetation
moisture
Bor Krai: discrimination of
Alisols and Acrisols
Huai Bong:
discrimination of
Alisols, Cambisols,
Regosols and Leptosols
SPOT 5
2 10
Visible (red):
0.61–0.68
Roads, bare soil;
discrimination of
vegetated/nonvegetated areas
All areas: discrimination of
Alisols, Acrisols,
Cambisols and
Technosols
3 10
Near Infrared:
0.78–0.89
Vegetation biomass,
water-vegetation
discrimination
Mae Sa Mai: discrimination
of Acrisols and
Cambisols
4 2.5
Panchromatic:
0.49–0.69
Provides higher
resolution
All areas: discrimination of
Alisols, Acrisols and
Cambisols
2 Beyond the Horizons: Challenges and Prospects for Soil Science and Soil. . .
61