An automatic method for land cover mapping and for detecting forest change has been designed for high resolution couples of image. The work is done on 20X20 km samples of 30 m resolution Landsat imagery. The methodology has been developed for two dates extracts but several couple of images can be compared. An automatic multi-date segmentation is applied on extracts pairs. Segmentation parameters are tuned, thanks to an iterative procedure, in order to provide image-objects of 1 hectare. In dense moist forest areas, 1 hectare guarantees a pure land cover for each object-segment. These image-objects will be the units of the classification and change detection work. An unsupervised classification is performed for each image, grouping image-objects into clusters. For each image, a tree cover mask is made based on a land cover map of reference with a coarser scale. Each cluster is then automatically labelled with this tree cover mask. Small image-objects are aggregated into image-object with a minimum mapping unit of 5 hectares thanks to a second segmentation level. Two independent things are then produced, a land cover map for each date of interest and a set of objects flagged as changed between the two dates. The forest change detection is obtained by running a statistical outlier detection method on the difference of both images. The detection is done under the union of the tree cover mask of both dates in order to work with a homogenous set of objects. The accuracy of this methodology is assessed for ten pairs of images, visually validated by experts.

VERHEGGEN Astrid;  ERNST Céline;  DEFOURNY Pierre;  BEUCHLE Rene'; 

2011-08-25

International Society for Photogrammetry and Remote Sensing (ISPRS)

JRC65179

http://www.isprs.org/proceedings/XXXVIII/4-C7/html/papers.html,    https://publications.jrc.ec.europa.eu/repository/handle/JRC65179,