In order to still improve classification, it is interesting to bring more information to the process of fusion. In addition to the spectral information provided by the spectral bands, information on the context of each class are added.

Out-image data have the same representation that spectral data: the samples of each class are used to build possibility distributions. Information thus arises in a numerical form like in the case of spectral information.

 Out-image data ``distance to roads'' and ``plateaux''

The out-image data ``distance to roads'' and ``plateaux'' give the better improvement of classification rates. A mean classification of 63.94% is reached against 53.35% with only the spectral bands.

The image obtained (table 20) presents more homogeneous areas for each class, but large areas of unclassified pixels appear. These two phenomena can be expressed by:

• The additional information brought by the out-image data ``distance to roads'' and ``plateaux'' are relevant and discriminating for the classes to recognize. Conjunctive fusion takes them into account correctly, because its behaviour reinforces concordant information.
  
• The very severe behaviour of conjunctive fusion and the nature of the data to be fused (scalar values instead of possibility distributions) result in a strong increasing of the number of unclassified pixels as the number of sources to be fused increases.
  
  In effect, all the sources must be at least a little in agreement about the classification of a pixel in a class c (i.e. give a degree of possibility of membership not null) so that a degree of membership not null is attributed to this class by fusion. If only one source among all the sources refuses to classify the pixel in class c, it is enough to discard this class from the set of the classes ``possible''.
  
  The mode of aggregation of conjunctive fusion is the selection of the plausible classes by the unanimity of the sources. It is then obvious that the larger the number of sources brought to come to a conclusion about the classification of a pixel is, the more difficult it is to reach this unanimity. On the other hand, this brings a reinforcement of the quality of the information available.

By adding information ``distance to roads'' and ``plateaux'' to the spectral bands, classification is improved of 10.59 points compared to classification with only the spectral bands (see table 20).

Classes 1, 3, 5 and 9 are better recognized :

• with an improvement of recognition of 52.94 points for class 1 (from 32.90% to 85.84%),
• of 25.17 points for class 3 (from 39.54% to 64.71%),
• of 35.73 points for class 5 (from 58.39% to 94.12%), and
• of 42.70 points for class 9 (from 43.14% to 85.84%).

The decreasing of recognition of class 2, 4, 7 and 8 is significant:

• 21.57 point for class 2 (from 50.11% to 28,54%),
• 10.23 points for class 4 (from 52.17% to 41.94%),
• 15.69 points for class 7 (from 70.37% with 54.68%), and
• 16.13 points for class 8 (from 64.71% to 48.58%)).

The final rate of classification of 63.94% does not correspond to a global improvement of classification, but to the best recognition of certain classes to the detriment of the others.

   

Unclassified pixels: 33.42%

Table 20: Conjunctive fusion of spectral bands MSS4 to MSS7 and of out-image : distance to roads and plateaux.

Rates of classification obtained
Class	Number of pixels        correctly classified     (A)	Number of pixels in samples          (B)	Rates of pixels      correctly classified    (A/B)
1	394	459	85.84%
2	131	459	28.54%
3	198	306	64.71%
4	164	391	41.94%
5	432	459	94.12%
6	313	459	68.19%
7	251	459	54.68%
8	223	459	48.58%
9	394	459	85.84%
TOTAL	2500	3910	63.94%

 Out-image data ``distance to roads'', ``plateaux'' and ``valleys''

The addition of a third out-image data, ``valleys'', does not improve the classification. The rate passes to 62.69% against 63.94% with two out-image data. The more there are sources of information to fuse, and the more it is difficult to obtain a area of agreement for each source.

So the rate of unclassified pixels still increases (from 33.42% of pixels not classified with six images to 49.98% with seven images!). It is the result of the severe behaviour of the conjunctive fusion which is unable to correctly manage the conflicts between a great number of sources. The limit of performance of conjunctive fusion is reached. Beyond, the results will be more and more degraded.

 Out-image data ``distance to roads'', ``plateaux'', ``valleys'' and ``slopes orientation''

The addition of a fourth out-image source, ``slopes orientation'', provides a decreasing of the average classification (from 62.69% for seven images to 59.51%). The conflict between the sources increases, involving the failure of the process of fusion in many cases and the none-classification of a great number of pixels (60.51%).

The degradation of the mean classification continues as sources to be fused are added. The fusion of the totality of the thirteen sources available leads to a very weak mean rate of classification of 22.89% (table 22) because almost all the pixels of the image (93.03%) are unclassified. The performance of conjunctive fusion is degraded by the problem of unclassified pixels. It is exactly the same problem with Bayes' rule. These methods of fusion are unsuited to fuse a great number of sources because they are unable to manage the conflict.

It is noted however that confusion between the classes largely decreases when the number of sources used increases. The reduction about the recognition of the classes is only due to the not classified pixels. All the pixels to be classified are located around the samples. The increase of the number of sources of information thus improves effectively the distinction of the classes. It remains to regulate the problem of the conflict between the sources so that classification becomes acceptable.

   

Unclassified pixels: 93.03%

Table 22: Conjunctive fusion of spectral bands MSS4 to MSS7 and of out-image data : distance to roads, plateaux, valleys, slopes orientation , slope steepness, crests, irrigation, DEM and distance to the urban areas

Rates of classification obtained
Class	Number of pixels        correctly classified     (A)	Number of pixels in samples          (B)	Rates of pixels      correctly classified    (A/B)
1	64	459	13.94%
2	88	459	19.17%
3	79	306	25.82%
4	44	391	11.25%
5	99	459	21.57%
6	185	459	40.31%
7	62	459	13.51%
8	37	459	8.06%
9	237	459	51.63%
TOTAL	895	3910	22.89%

Table 23: Matrix of inter-classes confusion in percentage due to classification using conjunctive fusion of spectral bands MSS4 to MSS7 and of out-image data: distance to roads, plateaux, valleys, slope orientation, slope steepness, crests, irrigation, DEM and distance to urban areas

	Classes observed
Class expected	1	2	3	4	5	6	7	8	9	Unclassified
1	13.94	-	-	-	-	-	-	-	-	86.06
2	-	19.17	-	-	-	-	-	-	-	80.83
3	-	-	25.82	0.33	-	-	-	-	-	73.86
4	-	-	-	11.25	-	-	-	-	-	88.75
5	-	-	-	1,09	21.57	-	-	-	-	77.34
6	3,49	-	-	-	-	40.31	-	-	-	56.21
7	-	-	-	-	-	-	13.51	-	-	86.49
8	-	-	-	-	-	-	-	8.06	-	91.94
9	-	-	-	-	-	-	-	-	51.63	48.37