Page 320 - Machine Learning for Subsurface Characterization

280 Machine learning for subsurface characterization

0.100 0.100
0.075 0.075
0.050 0.050
0.025 0.025
0.000 0.000
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
0 10 0 10
15 11 15 11
30 12 30 12
45 13 45 13
60 60
14 14
75 75
15 15
90 90
16 16
105 105
Sensors 17 Sensors 17
120 120
Source Source
135 18 135 18
150 28 150 28
0 15 30 45 60 75 90 105 120 135 150 0.00 0.05 0.10 0 15 30 45 60 75 90 105 120 135 150 0.00 0.05 0.10
0.000 Importance 0.000 Importance
0.025 0.025
0.050 0.050
0.075 0.075
0.100 0.100
19 20 21 22 23 24 25 26 27 28 19 20 21 22 23 24 25 26 27 28
Sensor index Sensor index
(A) (B)
0.100 0.100
0.075 0.075
0.050 0.050
0.025 0.025
0.000 0.000
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
0 10 0 10
15 11 15 11
30 12 30 12
45 13 45 13
60 60
14 14
75 75
15 15
90 90
16 16
105 105
Sensors 17 Sensors 17
120 120
Source Source
135 18 135 18
150 28 150 28
0 15 30 45 60 75 90 105 120 135 150 0.00 0.05 0.10 0 15 30 45 60 75 90 105 120 135 150 0.00 0.05 0.10
0.000 Importance 0.000 Importance
0.025 0.025
0.050 0.050
0.075 0.075
0.100 0.100
19 20 21 22 23 24 25 26 27 28 19 20 21 22 23 24 25 26 27 28
Sensor index Sensor index
(C) (D)
FIG. 9.24 Importance of sensor/receiver calculated using the feature permutation method for each
of the four datasets/experiments. The importance values are computed as the reduction in perfor-
mance of the trained voting classifier when feature values become non informative. (A) Experiment
#1. Four orientations, kappa ¼ 10, (B) Experiment #2. Four orientations, kappa ¼ 50, (C) Experi-
ment #3. Eight orientations, kappa ¼ 10, and (D) Experiment #4. Eight orientations, kappa ¼ 50.
travel time measured at various sensors/receivers. Classifiers can learn to detect
the differences in travel times and relate them to the various orientations of dis-
continuities. Interestingly, for Dataset #2, the sensor exactly opposite to the
transmitter is the most important, while those on the boundaries adjacent to
transmitter-bearing boundary are the least important (Fig. 9.24B). When devel-
oping classifiers for detecting eight classes/orientations (Fig. 9.24C and D), the
sensors on the adjacent boundaries are much more important than those required
for four-class classification (Fig. 9.24A and B). With increase in dispersion, the
sensors on the boundaries adjacent to the transmitter boundary become more
important.