Name
Abstract | ||
|---|---|---|
In this paper we explore the use of an ensemble of classifiers or multiple classifiers for classification of hyperspectral data. Traditionally, in pattern recognition, a single classifier is used to determine which class a given pattern belongs to. However, in many cases, the classification accuracy can be improved by using an ensemble of classifiers in the classification. In such cases, it is possible to have the individual classifiers support each other in making a decision. The aim is to determine an effective combination method that makes use of the benefits of each classifier but avoids the weaknesses. Three of the most used multiclassification approaches are boosting, bagging (1) and random forests (2). These approaches are based on manipulating training samples. In contrast, statistical consensus theory is based on treating data sources separately, and it uses all the training data only once. Furthermore, statistical consensus theory classifies the individual data sources, which can in the hyperspectral case be subsets of the hyperspectral data and combines the results using decision fusion. Decision fusion can be defined as the process of fusing information from several individual data sources after each data source has undergone a preliminary classification. For instance, Benediktsson and Kanellopoulos (3) proposed a multisource classifier based on a combination of several neural/statistical classifiers. The samples are first classified by two classifiers, every sample with agreeing results is assigned to the corresponding class. Where there is a conflict between the classifiers, a second neural network is used to classify the remaining samples. The main limitation of this method is the need of large training sets to train the different classifiers. In this paper, we consider using fuzzy decision fusion rules as proposed by Fauvel et al. (4) in order to aggregate the results of different classifiers, and conflicting situations, where the different classifiers disagree, are solved by estimating the pointwise accuracy and modeling the global reliability for each algorithm. Furthermore, we will use Support Vector Machines (SVM) to combine the contributions from individual sources as proposed in (5). Different airborne hyperspectral datasets are use for the experiments. A ROSIS-03 (Reflective Optics System Imaging Spectrometer) dataset acquired over the city of Pavia and an AVIRIS image from the region sorounding volcano Hekla, Iceland. The flight over the city of Pavia, Italy, was operated by the |
Year | DOI | Venue |
|---|---|---|
2008 | 10.1109/IGARSS.2008.4778793 | IGARSS |
Keywords | Field | DocType |
kernel,remote sensing,hyperspectral sensors,image classification,classification,hyperspectral imaging,maximum likelihood estimation,neural network,multidimensional systems,pattern recognition,support vector machine,training data,accuracy,random forest,imaging spectrometer,support vector machines,image processing,correlation,risk management,covariance matrix | Structured support vector machine,Kernel (linear algebra),Data mining,Pattern recognition,Computer science,Support vector machine,Hyperspectral imaging,Artificial intelligence,Margin classifier,Contextual image classification,Classifier (linguistics),Ensemble learning | Conference |
Citations | PageRank | References |
7 | 0.57 | 8 |
Authors | ||
6 | ||
Name | Order | Citations | PageRank |
|---|---|---|---|
| Jon Atli Benediktsson | 1 | 4064 | 251.17 |
| Xavier Ceamanos | 2 | 34 | 8.54 |
| Björn Waske | 3 | 435 | 24.75 |
| Jocelyn Chanussot | 4 | 4145 | 272.11 |
| Johannes R. Sveinsson | 5 | 1150 | 95.58 |
| Mathieu Fauvel | 6 | 742 | 42.30 |