104
for points when they are labelled manually and using autoencoder. The dominant geochemical assays for WA include SiO 2 , Fe, LOI and Al 2 O 3 , with median assays of 45wt%, 25wt%,
8wt% and 5wt% respectively. In contrast, NM is dominated by only Fe and SiO 2 , with medians of 59wt% and 25wt% respectively. There is also a small amount of LOI. The different WA
groups have similar assays (Figures 1 (a) and 1(b)), with the only significant change occurring in the top quartile of the silica. While the overall medians have also stayed the same for
the NM, the autoencoder results have more outliers, particularly in the Fe and Al 2 O 3 . The
misclassifications tended to occur at the base of the WA. This section contains less shale and
more iron ore than the rest of the member, and is sometimes more similar to the NM than
the rest of the WA.
The accuracy of Test 3 (Table 3) was higher than that of Test 1 (Table 1) for all methods.
This indicates that the geochemical assays contain more information that is useful for separating the WA and NM. However, the accuracy of Test 3 (Table 3) was lower than Test 2
(Table 2), which indicates that the geochemical assays are a better indicator of rock type than
lithology.
For all tests, autoencoder and SVM had significantly better results than k-means. This is
due to k-means not using training data. Although they had similar results, autoencoder is
preferable to SVM for this application as it can handle more than two categories. Therefore
for this problem is the best of the tested methods.
5 CONCLUSION
Deep learning using autoencoder, SVN and k-means were successfully applied to both geochemical assays and mineral groups to determine the rock type or lithology. Autoencoder
Table 3. Results from Test 3, using geochemical assays to identify lithology
Method
Accuracy (%)
Sensitivity (%)
Specificity (%)
Autoencoder
88.4
89.6
88.5
K-means
31.9
39.5
22.6
SVM
88.5
93.2
86.6
Figure 1. The distribution of the geochemical assays of: (a) Points manually labelled as WA.
(b) Points labelled by the autoencoder as WA. (c) Points manually labelled as NM. and (d) Points
labelled by the autoencoder as NM.
80
70
-:70
60
60
~
-:u
50
50
40
g
40
g
30
30
20
20
T
8
~ 1
10
_;_
8
~
.
10
_;_
_;_
1
~
c::::,
c::::,
.......
a)
Fe
At203
5102
LOI
MN
MGO
K20 TI02
b)
Fe
Al203
5102
LOI
MN
MGO
K20
TI02
70
-:80
60
8
70
_;_
60
6
50
_;_
50
40
.:.
40
30 $
30 $
20
~
20
10
10
.....
--<-_..._
~
c)
Fe
A1203
5102
LOI
MN
MGO
K20
TI02
d)
Fe
Al203
5 102
LOI
MN
MGO
K20
TI02
for points when they are labelled manually and using autoencoder. The dominant geochemical assays for WA include SiO 2 , Fe, LOI and Al 2 O 3 , with median assays of 45wt%, 25wt%,
8wt% and 5wt% respectively. In contrast, NM is dominated by only Fe and SiO 2 , with medians of 59wt% and 25wt% respectively. There is also a small amount of LOI. The different WA
groups have similar assays (Figures 1 (a) and 1(b)), with the only significant change occurring in the top quartile of the silica. While the overall medians have also stayed the same for
the NM, the autoencoder results have more outliers, particularly in the Fe and Al 2 O 3 . The
misclassifications tended to occur at the base of the WA. This section contains less shale and
more iron ore than the rest of the member, and is sometimes more similar to the NM than
the rest of the WA.
The accuracy of Test 3 (Table 3) was higher than that of Test 1 (Table 1) for all methods.
This indicates that the geochemical assays contain more information that is useful for separating the WA and NM. However, the accuracy of Test 3 (Table 3) was lower than Test 2
(Table 2), which indicates that the geochemical assays are a better indicator of rock type than
lithology.
For all tests, autoencoder and SVM had significantly better results than k-means. This is
due to k-means not using training data. Although they had similar results, autoencoder is
preferable to SVM for this application as it can handle more than two categories. Therefore
for this problem is the best of the tested methods.
5 CONCLUSION
Deep learning using autoencoder, SVN and k-means were successfully applied to both geochemical assays and mineral groups to determine the rock type or lithology. Autoencoder
Table 3. Results from Test 3, using geochemical assays to identify lithology
Method
Accuracy (%)
Sensitivity (%)
Specificity (%)
Autoencoder
88.4
89.6
88.5
K-means
31.9
39.5
22.6
SVM
88.5
93.2
86.6
Figure 1. The distribution of the geochemical assays of: (a) Points manually labelled as WA.
(b) Points labelled by the autoencoder as WA. (c) Points manually labelled as NM. and (d) Points
labelled by the autoencoder as NM.
80
70
-:70
60
60
~
-:u
50
50
40
g
40
g
30
30
20
20
T
8
~ 1
10
_;_
8
~
.
10
_;_
_;_
1
~
c::::,
c::::,
.......
a)
Fe
At203
5102
LOI
MN
MGO
K20 TI02
b)
Fe
Al203
5102
LOI
MN
MGO
K20
TI02
70
-:80
60
8
70
_;_
60
6
50
_;_
50
40
.:.
40
30 $
30 $
20
~
20
10
10
.....
--<-_..._
~
c)
Fe
A1203
5102
LOI
MN
MGO
K20
TI02
d)
Fe
Al203
5 102
LOI
MN
MGO
K20
TI02