Automating woody vegetation change detection at regional scales: the problem of clouds and cloud shadows

This paper presents a semi-automated cloud and cloud shadow masking method, developed for SPOT5 satellite data over New South Wales (NSW), Australia. As clouds are very similar to several surface features in these data, attempts at using traditional image classification techniques have not been effective. The new method uses morphological feature extraction, where marker pixels are identified using image-specific, automatically defined criteria, and mask segments are grown using the watershed from markers transformation. Manual input is only required to add and delete marker points in order to improve the masks. The method has achieved an overall accuracy of around 80 %, with most errors due to the commission rather than omission of pixels. This has proved satisfactory, and the method has been incorporated into the processing stream for woody vegetation change detection by the NSW government.