Name
Abstract | ||
|---|---|---|
Land cover classification has remained one of the central research foci in remote sensing for the last decades. Effective tools that support the inventory of land cover changes are of paramount importance to analyze environmental and natural impacts on the biosphere. During the past decade the remote sensing research community has shown a growing interest in Ensemble Classifier Systems, which are also known as multiple classifier systems, committee of classifiers, or mixture of experts. These approaches provide several advantages compared to traditional monolithic or single expert classifier designs. This study addresses the problem of supervised land-cover classification demonstrated on the Iberian Peninsula, using hypertemporal data from the "Mediterranean Extended Daily One Km AVHRR Data Set" (MEDOKADS). The data set is compiled from the NOAA-11, -14 and -16 sensors and is distributed by the Institute of Meteorology, Free University of Berlin (1). 10 day NDVI maximum value composites from each individual year within the 1989 to 2004 period were preprocessed using Minimum Noise Fraction (MNF-) transformation. 11 aggregated land-cover classes derived from the CORINE 2000 land-cover map (Co-ordination of Information on the Environment) (2) were selected to train the classifiers and to validate results. The composition of the training and test data sets was adjusted according to the abundances of each land-cover class in the study area: 25% of the available pixels of each class (but not more than 5000 samples) were randomly selected to define the data sets. The MNF-scores were then individually pre-classified for each year using support-vector machines (SVM)(3). The continuous outputs from the SVM, which can be interpreted in terms of posterior probabilities, where then used in the decision fusion approach to train a second-order SVM classifier to merge the information from the individual classifiers (4). The considered periods for the approach varied between two and five years. The results for the fusion strategy were significantly better compared to the annual classifications. Extending the period from two year to the five year period also improved the accuracy. The decision fusion approach was compared with another ensemble method (majority voting) and with a single expert classifier that was trained using all MNFs from the considered multi-annual periods. Regarding classification accuracy and the balance between errors of omission and commission, data fusion was superior to the other approaches (figure 1). Furthermore, this method was characterized by the smallest classification variability. The results suggest that decision fusion is a universal method to merge information derived from multisensor as well as from singles sensor systems, as long as the information in the pre-classification steps is different for each classifier. |
Year | DOI | Venue |
|---|---|---|
2008 | 10.1109/IGARSS.2008.4779524 | IGARSS |
Keywords | Field | DocType |
vegetation,posterior probability,probability density function,spatial resolution,classification algorithms,data fusion,sensor fusion,second order,support vector machines,multispectral imaging,support vector machine,temperature,data mining,remote sensing,majority voting,accuracy,voting,ndvi,image classification | Pattern recognition,Computer science,Remote sensing,Support vector machine,Sensor fusion,Posterior probability,Normalized Difference Vegetation Index,Artificial intelligence,Statistical classification,Contextual image classification,Majority rule,Land cover | Conference |
Citations | PageRank | References |
0 | 0.34 | 6 |
Authors | ||
4 | ||
Name | Order | Citations | PageRank |
|---|---|---|---|
| Thomas Udelhoven | 1 | 47 | 11.56 |
| Björn Waske | 2 | 435 | 24.75 |
| Sebastian van der Linden | 3 | 84 | 10.59 |
| Sonia Heitz | 4 | 1 | 0.70 |