Documentation and Recovery of Rupestrian Paintings: An Automatic Approach
379
4. SPECTRAL CLASSIFICATION
Several supervised classifications were performed taking into account multispectral and multiband images, although herein only
appears one of the best multiband classifications (Puertas, 2000). Training sets were collected for the seven features from well-
distributed and separately placed polygons all over the cave. A class for the supervised classification procedure was assigned to each
feature.
The classifier used in this study was the maximum-likelihood with a 99% confidence level threshold. Therefore, after classification
the non-classified pixels were assigned as Null class. This item was not considered by itself in the error matrix form for the accuracy
assessment. The resulting image appears in Fig. 3.
Legend
□ Pigment
m Reddish stone
^ Brownish stone
Rosy stone
m Oxide
D Whitish stone
□ Darkish stone
| Null class
□
Fig. 3: Maximum-likelihood classified image taking into account seven bands
5. RESULTS AND DISCUSSION
After the pixel assignment was completed, the accuracy assessment within the entire study site was carried out by means of a set of
test pixels and a set of classified pixels. The results were presented in error matrix form (Table 1); the producer's accuracy and user's
accuracy were computed afterwards (Table 2).
Table 1: Error matrix form for accuracy assessment.
Class
Oxide
W.s.
Rosy s.
Red. s.
Br. s.
D.s.
Pig.
Total
Oxide
962
0
8
3
42
0
0
1015
Whitish stone
0
2552
1
2
7
0
1
2563
Rosy stone
31
2
1066
28
25
0
0
1152
Reddish stone
0
0
6
946
2
0
54
1005
Brownish stone
35
12
48
15
1023
0
24
1157
Darkish stone
5
0
0
7
5
1233
239
1489
Pigment
0
2
2
127
10
3
1644
2088
Total
1033
2568
1126
1128
1114
1236
2262
10469
Table 2: Producer's accuracy and user's accuracy.
Class
Producer's accuracy
(%)
User's accuracy (%)
Oxide
99.1
94.8
Whitish stone
99.4
99.6
Rosy stone
94.5
92.5
Reddish stone
83.9
94.1
Brownish stone
91.8
88.4
Darkish stone
99.8
82.8
Pigment
85.9
93.1
