Recent Advances in Remote Sensing Time Series Image Classification

This chapter revises the recent advances in the automatic classification of remote sensing (RS) image time series (images regularly acquired by satellite-borne sensors on the same areas at different times) to update land-cover maps. The availability of up-to-date land-cover maps has significant impacts on several aspects of daily life from the economical, administrative and management point of view. Thus land-cover maps updating by classification of RS images is an hot topic. This is even more true since an increasing number of image time series are being acquired and freely available for a large number of satellite missions (e.g., Landsat archive, ESA Sentinel missions). Land-cover maps can be updated by direct supervised classification of each image in the time series. However, in order to properly train the classifier such an approach requires reliable ground reference data for each available temporal image. In operational scenarios, gathering a suffcient number of labeled training samples for each single image to be classified is not realistic due to the high cost and the related time consuming process of this task. To overcome these problems, domain adaptation (DA) methods have been recently intro- duced in the RS literature. Accordingly, in this chapter the most recent methodological developments related to DA are presented and discussed by focusing on semi-supervised and active learning approaches. Finally, the most promising methods to define low-cost training sets and to effectively classify RS image time series will be discussed.