A classification procedure is developed that distinguishes between pix
els that are clearly associated with a given class versus those where
class assignment is uncertain. Alternatively, most commonly used class
ification algorithms force each pixel into a single class without rega
rd to certainty. By classifying pixels in order of certainty and consi
dering spatial context, pixels with weak observational evidence for cl
assification are prevented from contributing to their neighbor's decis
ions. Subsequently, a better decision is made for the uncertain pixels
by considering the previously classified neighbors. Degrees of certai
nty measures can assist in later map accuracy assessment by allowing f
or stratified sampling nf zones having similar certainty levels.