Selection of samples for active labeling in semi-supervised hyperspectral pixel classification

摘要

One of the problems in semi-supervised land classification tasks lies in improving classification results without increasing the number of pixels to be labeled. This would be possible if, instead of increasing the amount of data we increased the reliability of the data. We suggest to replace the random selection by a unsupervised clustering based selection strategy in building the training data. We use a mode seeking clustering method to search for cluster representatives, which will be labeled and then used for training. Here an improvement to the result of the clustering algorithm is introduced by taking advantage of the spatial information in the image. The number of selected samples provided by the clustering can be reduced by using a spatial-density criterion to dismiss redundant training information. Two different alternatives are considered for a spatial criterion, one dismisses selected samples in the same neighbourhood and the other includes the pixel coordinates for giving the spatial information a larger weight in the clustering. Both alternatives improve the classification-segmentation results. The classification scheme with training selection provides state-of-the-art pixel classification results using a smaller training set and suggests an alternative to random selection.