Remote Sensing

Image Classification

Image classification is performed by the remote sensing software and it requires multispectral image having at least two channels. Classification is based on the spectral signature of an object. Any pixel in the image has a set of values (gray shades, digital numbers, etc.) and each value corresponds to a different spectral band. This set of values represents the spectral signature of the object. The idea of classification is based on the fact that each object has a unique spectral signature, which theoretically is correct but requires many spectral channels (high spectral resolution) to obtain this in practice. The regular multispectral images have a limited number of channels i.e. the SPOT/XS image has three channels, the Landsat-TM image has seven channels etc. There are images, however, called "Hyperspectral" which could have as many as 1024 channels. For images of limited spectral resolution only a relative small number of classes of objects can be identified as having quite distinct spectral signatures. The spectral signature, on the other hand, is not just a set of numerical values. It is in fact a set of average values accompanied by their variances. Computed covariances between different class spectral signatures indicate the degree of correlation of such classes and help to select classes with distinct spectral signatures.

Image classification algorithms try to create groups of pixels of similar spectral signatures and the degree of such similarity is defined by certain criteria. The groups of pixels are called clusters and they are formed if the vector of values defining the spectral signature of a pixel is considered as a vector of coordinates in the spectral space. Plotting all pixels in the spectral space results to the formation of such clusters. Clusters are not often clearly separated from one another and there are discrimination algorithms based on certain criteria, as mentioned above, which draw the borderlines among clusters. It is interesting to notice that the spectral space is defined by the number of spectral bands in the image i.e. two spectral bands define a two dimensional spectral space while N-spectral bands define an N-dimensional spectral space. In the two dimensional spectral space clusters have the shape of an ellipse, in the three dimensional spectral space clusters have the shape of an ellipsoid, and in the N-dimensional spectral space clusters have the shape of a hyperellipsoid.

There are two kinds of classification:

1. Supervised classification, and
2. Unsupervised classification.

Supervised classification is based on spectral signatures of known classes in the image. The areas of such classes are identified in the image and each class category is precisely determined by ground truth or ground samples. Ground samples are obtained either before the image is taken and then the sampling areas are precisely transferred in the image, or, the image is obtained first then sampling areas are defined in the image and the class category in such area is determined by ground surveys. More often, however, we use ground truth that exists before taking the image (usually coming from existing thematic maps) and if it is necessary to obtain new ground samples we go in the field and obtain them. Minimum sample size is important to cover an area of about 40 – 100 pixels. Samples in supervised classification cover about 0.3% - 0.5% of the total area of the image. Samples may also be taken from interpretation of aerial photographs.

Unsupervised classification is performed by the software and it is based on spectral signatures of a number of classes in the image determined by corresponding clusters. Clustering algorithms can create clusters in the spectral space of any given number. The operator usually selects the clusters with good separation from the others. Statistical methods help to define such clusters. Unsupervised classification helps to determine the number of classes which are clearly separated and works as a planning tool to define areas of sampling for supervised classification.

Following is a complete classification example illustrating the basic steps of the process:

A Landsat TM image is composed of 10-lines by 10-columns. The following 3-bit dynamic range digital values are recorded in a file and they have the following order: 10-values from channel 3 (?=0.63 - ?=0,69), 10-values from channel-4 (?=0,76 - ?=0.90), 10-values from channel 3, 10-values from channel-4, …, etc.

1,2,0,1,2,5,6,5,6,7,7,7,6,6,6,7,7,6,5,6,1,0,1,3,4,7,5,6,7,7,6,6,5,5,5,6,5,6,5,7

0,2,1,2,3,6,5,5,6,7,5,6,6,4,4,6,6,4,4,4,1,3,4,3,2,4,5,6,7,6,6,4,4,4,3,4,3,3,3,3

3,3,3,3,3,2,7,0,1,2,4,3,3,5,4,4,3,2,2,2,2,4,6,6,5,5,6,0,1,0,4,4,4,3,2,3,3,1,1,1

6,6,6,5,7,6,6,2,2,1,3,2,3,3,2,2,4,2,1,1,7,6,5,1,1,0,2,1,0,1,2,2,4,2,1,1,1,1,0,2

7,2,1,1,2,2,1,2,2,0,3,1,1,0,0,0,0,2,1,1,1,0,0,2,1,1,1,0,1,0,2,0,1,1,0,1,0,0,1,2

Do the following:

1. Assemble the image of each channel
2. Create the clustering diagram (plot all pixels in the spectral space) and determine the following class categories using the minimum distance criterion:

1. Broad leaf vegetation
2. Sea water
3. Pasture
4. Olive grows
5. Urban environment

Notice that broad leaf vegetation has high reflectance in the near infrared (channel-4) and low reflectance in the red (channel-3), Sea water has low reflectance in both channels, Pasture has high reflection in channel-3 and lower than the vegetation reflection in channel-4, Olive grows are between the broad leaf vegetation and pasture, and the urban environment has high reflectance values in both channels.

1. Determine the spectral signature of each category by computing the average value from each channel.
2. Plot the classification results in a land cover thematic map form.
3. Assuming that the spatial resolution of the image is 150 m, compute the area for each class category.

PROBLEM SOLUTION

(1) Assembled images

(2) Clustering diagram

The number in a table cell indicates the accumulated pixels in that location

(3) Spectral signature of each class category

(4) Thematic map with class categories

(5) Estimation of areas

© 1999 TRIANET, Program of the European Union Socrates-Comenius
Last update on 04.05.1999 by Markus Zapke-Gründemann