Thresholding makes it possible to select the pixels according to their radiometry. The context of the pixels is not taken into account and it is necessary to have images with a good contrast and a good homogeneity.

As noticed by T LIANGHU [Lianghu, 1992], the use of multispectral images makes it possible to improve the results but they remain partial. The determination of the threshold is delicate and the value is different for each set of images. This method can however constitute a complement to a more elaborate processing.

Thresholding can be carried out

• on image composing multispectral information,
• on néo-channels combining the XS components of SPOT,
• or using different representation, for example HSI.

HAALA and VOSSELMAN [N. Haala, 1992] use channel H, deduced from RGB images, to detect roads which appear here in over brightness. A region growing, followed by a simple thresholding is sufficient for isolating roads. Rivers are extracted by thresholding from channel R. It appears in this case that the choice of the thresholds is not critical.