Automating image classification, a preliminary foray using archival GIS data to label pixels

This paper proposes a procedure to update existing mapping from a classified image in the Mary River floodplain of the Northern Territory, Australia. This procedure explicitly rests on the assumption that the spatial distribution of land cover types depicted in the reference cover type mapping is "still mostly correct", (and contains sufficient still-valid information to train a classification of a more recent image), with a view to updating the cover type mapping. Empirical probabilities of spectral class to cover type were estimated from GIS overlay of a classified TM image over existing mapping. Cover types are additionally characterised as mosaics of characteristic (template) textures. The image texture of a window around each pixel was estimated and compared for similarity to the GIS derived cover type texture templates. Pixels were labelled to the most 'likely' cover type. The 'likelihood' of a given cover type label was estimated by combining both the similarity of pixel neighbourhood texture and the the pixel spectral class probability of association to that cover type. The resulting map does not conform completely to the original mapping; however enough cover types do map to plausible locations to warrant continuing the investigation