Automatic Selection of Training Samples for Multitemporal Image Classification

The present work presents and evaluates a method to automatically select training samples of medium resolution satellite images within a supervised object oriented classification procedure. The method first takes a pair of images of the same area acquired in different dates and segments them in homogeneous regions on both images. Then a change detection algorithm takes stable segments as training samples. In experiments using Landsat images of an area in Southwest Brazil taken at three consecutive years the performance of the proposed method was close to the performance associated to the manual selection of training samples.