Uncertainty is an intrinsic feature of remotely-sensed image data for several reasons, including:

Resolution: the resolution of a remotely-sensed image defines the area represented by a single pixel. If a pixel represents a 100m by 100m region it is almost certain that its contents will be of mixed landcover types, and hence will generate data that reflects this mixture. Specific techniques (generally based on soft classifiers such as Idrisi’s UNMIX, BAYCLASS and FUZCLASS functions) are available for analysing within-cell spectral variation, often described as sub-pixel classification

Temporal variables: time of day and season affect both the landcover (e.g. growth stages) and illumination patterns (e.g. lighting of slopes, vegetation surfaces, moisture levels) resulting in additional variations within and between image datasets and bands

Spectral response: remote-sensing may not measure the spectral response band or bands on which the particular surface provides maximum response. The response data may therefore be weak or mixed/easily confused with other surface responses

Soft classification procedures can provide an image displaying the degree of uncertainty of the classification. For example, for every soft classification image one pixel or set of contiguous pixels (region) may have the same or very similar degree of class membership, hence no one class is clearly favoured. This may be contrasted with a hard classification) process in which the winning class, i.e. the one with slightly higher probability or degree of membership, would be selected in preference to any other. An uncertainty map illustrates the degree to which one class (from a total of P classes) stands out from the rest for each pixel in the image. Definition of uncertainty could be based on Shannon’s information statistic (entropy) statistic or variations thereof, or on an application-specific measure. Idrisi adopts the latter approach, using an uncertainty measure, U, of the form:

Here, m is the maximum degree of membership recorded for the pixel in question, and  is the sum of the degree of membership for that pixel. Thus, with P=10 classes and m=0.3 and s=1, say, we have U=1‑0.22=0.78 whilst with m=0.1 we have U=1 (complete uncertainty) and with m=1 U=0 (complete certainty).