Semi-Supervised Classifier Ensembles for Classifying Remote Sensing Data

The analysis of data sets, which were acquired within different time periods over the same geographical region is interesting for updating land cover maps and operational monitoring systems. In this context an adequate and temporally stable classification approach is worthwhile. In the presented study a classifier ensemble (i.e., random forests) is trained on a multispectral image from an agricultural region from and is successively modified and adapted, to classify a data set from another year. A detailed accuracy assessment clearly demonstrates that the proposed modification of the classifier significantly improves the overall accuracy, whereas a simple transfer of a classifier to a data set from another year is limited and results in a decreased accuracy. Thus the proposed approach can be recommended for classifying multiannual data sets and updating land cover maps.