MONO-BAND CLASSIFICATION

MONO-BAND CLASSIFICATION

Classification consists in assigning each pixel to a class, or assuming that it does not belong to any proposed classes and to store it in a class of rejection. The decision to rather assigning a pixel to a class from other is performed using the degree of membership of the pixel. The choice of the target class comes to find the class having the highest degree of membership.

For each pixel x of each image I_i , a degree d_i^c(x) of membership of each class c , is calculated by means of a probability or possibility measure. Fusion must, for a given pixel x, aggregate information about its belonging to some classes.

For this pixel x, information is represented by p vectors I (p information sources) of dimension k (k classes). Information fusion aggregates the measures contained in these p vectors onto a sole final vector I (figure 15). The pixel is then stored in the class having the highest degree of membership.

**Figure 15:** Information fusion with a pixel x.
$\begin{figure}\begin{center}\epsfbox{c4-fusion.eps}\end{center}\end{figure}$

To highlight that a correct classification can not be carried out by using one spectral band only, classification is achieved for each spectral band. I.e. classification of a spectral band I_b , takes only into account the information contained in this one.

The reasons for this impossibility are the lack of discrimination between each class, the impurity of the information brought by each pixel, and the noise of sensors spoiling the available information.

Classification rates

Tables 5 to 8 give the rates of classification obtained by using a Bayes classification for each of the four spectral bands and the corresponding resulting images.

Those rates are lower when fusion is performed with the operator "minimum" of possibility theory.

It is clear that these isolated classifications lead to poor results:

at best, 40.90% of the pixels on average are correctly classified (spectral band MSS4);
in the worst case, this rate falls to 24.91% for band MSS7.

There are different reasons for bad classification, for instance confusion between classes and noise due to sensors.

Confusion between classes

Those bad rates are due to the confusion made between classes. As their spectrum are very similar, a weak grey level change is enough to classify a pixel from one class to another.

Confusion matrices of spectral band MSS4 to MSS7 show that class 2 is often merged

with class 4 for the spectral bands MSS4 and MSS5 (in 55.24% of case for MSS4, and in 50.38% of case for MSS5);
rather with class 6 for band MSS6 (in 27.89% of case)
rather with class 5 for band MSS7 (in 33.77% of case). .

Classes which histogram has a narrow spread out width, and well marked peak, are recognized best (such classes 1 and 2 for example), whereas classes with a more spread out histogram are merged with the preceding ones and are less recognized.

Noise

Another cause of these bad results is due to the noise added by the sensors to the images.

It is particularly visible in spectral band MSS6. The basic image contains a frame. This one is found in the corresponding classification (table 7). That shows that the defect of the sensor of MSS6 involves errors of classification.
Band MSS4 allows to recognize class 1 rather well (low forest sempervirente), 6 (mosaic of shrubby savanna and deforestations), 7 (irrigated cultures), and 8 (cultures dry). On the other hand it distinguishes very badly classes 3 (semi-decidue vallicole forest), and 9 (discontinuous thickets and pseudo-steppe with rock exposures).

Band MSS5 allows to recognize rather classes 7 (irrigated cultures), 8 (cultures dry), and 9 (discontinuous thickets and pseudo-steppe with rock exposures). For this band there is not a class really little recognized.

Band MSS6 only allows to recognize sufficiently two classes. They are classes 5 (raised savanna) and 8 (cultures dry). Class 6 (mosaic of shrubby savanna and deforestations) is rather slightly recognized.

In band MSS7, the classes are rather badly recognized, excepted classes 2 (semi-decidue wet dense forest) and 7 (irrigated cultures). Histograms representative of classes 8 (dry cultures) and 9 (discontinuous thickets and pseudo-steppe with rock exposures) are always outclassed by those of classes 5 and 7. Class 4 (semi-deciduous forest or dry deciduous) is very badly recognized.

At last, classes 7 and 8 are generally the best recognized (each one is well recognized by three spectral bands), then classes 1, 2, 5, 6 and 9.

The less recognized classes are classes 8 and 9, and, to a lesser extent, classes 3, 4 and 6.

The conversion of the histograms into possibility distributions does not really improve the discernability between sources. It is thus necessary to resort to other techniques to improve classification. The fusion of data will make it possible to raise, at least better, ambiguities related to classification.

Classification Rates

Confusion Matrices

IRIT-UPS