- 328 -
2. PROPERTIES OF LANDSAT DATA
A pixel in a LANDSAT image is a 4-dimensional vector formed by four
integer variables. Each one of these components is quantized into 64 or
128 possible values. The total number of possible vectors is 24^, i.e.,
over a hundred million. In fact the number of different vectors that one
finds in a LANDSAT picture is much smaller, of the order of a few thousand
[1]. This is due to the high correlation of spectral intensities in the
four multispectral scanner bands. It has also been found {21 that the high
correlation means that one can find a 2-dimensional plane in the 4-dimensional
space which is approximately parallel to all the vectors in a given image.
This is expressed by saying that MSS data is essentially two-dimensional.
Four our study we have selected a subimage of 512 lines and 470 columns
covering an area which is well known by previous studies [3]. To study the
correlation of spectral band we have computed the covariance matrix and the
correlation matrix which appear in Table 1. We have also computed the
eigenvalues of the correlation matrix which are the following 2.94, 0.96,
0.09, 0.01. This means that the first two principal components contain 97.57
of the information contained in all the picture. Nevertheless for reasons
of simplicity we have chosen to work with two of the natural bands, namely
bands 4 and 7 which are the less correlated. Finally we compued the
2-dimensional histogram of bands 4 and 7 which shows only 1279 different
vectors. The preparation for the classification includes the definition of
training fields for seven classes, as show in Table 2.
3. TABLE CLASSIFICATION
The method for classification uses a table which is a matrix of 256
times 256 elements. Each element in the matrix is a byte containing a number
or symbol that represents a class. Therefore the maximum number of classes
that can be considered is 256. The two indices of the matrix will be two
variables which can be the spectral intensities of two of the four MSS bands
or, in general, two linear combinations or two arbitrary functions of the
four bands, resulting from a coordinate transformation, for example the
principal component transformation. The only restriction to that transformation
will be that the resulting variables become quantized in the range 0-255.
The CPU time needed to perform a classification with this table in an
IBM 370/155 computer is of the order of 6 microseconds per pixel. Therefore
the classification of a full LANDSAT image containing 7,6 million pixels
can be completed in less than a minute of CPU time. This is to be compared
with the time taken by the multivariate normal maximum likelihood classifier,
currently used in the ERMAN-II system. This classifier computes the likelihood
for every pixel and class. This computation needs F(F-3)/2 multiplications,
F being the number of bands taken into account (usually, four). Estimating
the time for multiplication in 7 microseconds for the 370/155 we get for
the 7 class problem using 4 bands a CPU time of 686 microseconds per pixel,
this is over two orders of magnitude higher.
The previous considerations show that the actual process of'classification
with a given table is a very fast and straightforward task. We must now
consider the task of generating a table. If we still want to use the maximum
likelihood rule we only need to classify the 256 x 256 possible vectors, thus
generating the table. Of course this way of proceeding has only sense if the
final number of pixels to be classified is larger than 256 x 256. In the