In his article, T LIANGHU [Lianghu, 1992] presents a contextual method which simulates a visual perception, by using attributes of form, length and narrowness, as well as multispectral information.

He points out that most simple operators are sensitive to noise and work only on images with one spectral band. In such images, it is frequent to see the lineaments merging with their environment, whereas multispectral images make it possible to distinguish from the differences in albedo according to the wavelength.

Note that the albedo also called ``reflectivity'' is the ratio of the outgoing solar radiation reflected by an object to the incoming solar radiation incident upon it.

The combination of multispectral information and information space, contextual, improves quality of the extraction. The process is the following:

• First phase  of the process carries out a segmentation based on a classification by spectral band. Statistics on samples give the values of average and standard deviation for each band. Then, the pixels are classified independently.
  
  
• Second phase  eliminates isolated pixels due to the noise. For each pixel, the 8 neighbours are used to calculate four elongations according to the four directions: horizontal, vertical and two diagonals. If one of the elongations is higher than a preset threshold, the pixel is regarded as an element of the lineament.
  
  
• Third phase fills occlusions of small length 1 to 2 pixels by analysing the distribution of the elements of the lineaments in windows of size 3x3 and 5x5.
  
  

Lineaments are mostly identified in a good way, but often strongly fragmented. This is due to the following phenomena:

• spectral classification is disturbed by the variations of albedo of the lineament environment, and
• an excessive curve of the lineament does not make it possible to calculate the elongation correctly.

This process shows the utility of the multispectral data, but the selection of the pixels belonging to the lineaments is only very partial.