Land-Cover Classification of Hypertemporal Data using Ensemble Systems

This study addresses the problem of multiannual supervised land-cover classification using hypertemporal data from the "Mediterranean Extended Daily One Km AVHRR Data Set" (MEDOKADS) and a decision fusion approach. 10 day NDVI maximum value composite data from the Iberian Peninsula for every year in the observation period (1989 to 2004) were preprocessed using Minimum Noise Fraction (MNF-) transformation. The MNF-scores from each year were then individually pre-classified using support-vector machines (SVM). The continuous outputs from the SVM, which can be interpreted in terms of posterior probabilities, where used to train a second-order SVM classifier to merge the information within consecutive years. The decision fusion strategy significantly increased the classification accuracy compared to pre-classification results. Increasing the temporal range in decision fusion from a two year to five-year period enhanced the total accuracy. The outcomes from the selected approach were compared with another ensemble method (majority voting) and with a single SVM expert that was trained for comparable multiannual periods. The results suggest that decision fusion is superior to the other methods.