Estimation of the Areal Extent of Land Cover Classes that Only Occur at a Sub-Pixel Level

Mixed pixels often dominate coarse spatial resolution remotely sensed data used widely in land cover mapping at regional to global scales. Unfortunately, mixed pixels cannot be appropriately accommodated by conventional image classification techniques and consequently the spatial representation of land cover classes and estimates of their extent derived from such classifications may be erroneous. Unmixing the class composition of pixels allows the land cover distribution to be depicted in fraction images from which more accurate estimates of class extent may be derived. Problems with unmixing methods limit their application and here an alternative approach, based on an artificial neural network is used. The method was applied in unmixing the class composition of AVHRR data of Brazil, with emphasis on the accuracy with which the extent of classes that always occurred at a sub-pixel level in the data set could be estimated. The results showed a close correspondence between the predicted and actual extent of classes in AVHRR pixels (correlation coefficients ranged from 0.30–0.74, all significant at the 99% level of confidence). Furthermore, the estimated extent of the classes over the test site was generally much closer to the actual extent of the classes than estimates derived from a conventional classification based approach.