Object-based method for automatic forest change detection

A new method has been developed in order to automatically detect land cover changes in forested areas on a multitemporal dataset. From a multitemporal segmentation on the calibrated reflectance of all images, unchanged but especially the changed stands are accurately delineated. Stands are characterized by features extracted from the reflectance difference images. As these features for the changed objects will appear as outliers with respect to the ones for unchanged objects, they are identified through a multivariate iterative trimming procedure. The method, which was tested in eastern Belgian forest using three SPOT HRV images covering a decade, could detect accurately clearcuts and regenerations on both coniferous and hardwood. The performance of this method of change detection, measured by the detection accuracy, was proved to be higher (85 to 95 %) than a particular multidate classification, named RGB-NDVI (49 to 65%). The originality of this study is (i) the fact that an object-based approach is used instead of the classical pixel-based methods, and (ii) the automation of the process.