Ensemble Methods For Spectral-Spatial Classification Of Urban Hyperspectral Data

During the last twenty years hyperspectral data has become increasingly available. In land cover classification of remote sensing data, it is desirable to use spectral-spatial data in order to extract as much information as possible about the area being classified. The use of mathematical morphology for spectra-spatial classifications is one example, which was extensively discussed on context of the classification of urban remote sensing imagery. Often a range of morphological operators (i.e., opening and closing by reconstruction) with a structuring element of increasing size is applied to the data set and a so called morphological profile (MP) is generated. Nevertheless, the use of several morphological operators often results in high-dimensional data sets, which generally contain redundant and irrelevant information. Thus, adequate classification schemes are necessary.In many remote sensing studies classifier ensembles have been successfully used to handle high-dimensional data sets. In contrast to standard classifiers, which based on only one classifier decision, the approach combines several, but different classifier outputs. In doing so the overall accuracy is usually increased. Random Forests (RF) are one example of such a classifier system.The spatial information is derived by mathematical morphology and principal components of the hyperspectral data set, generating a set of different morphological profiles. The whole data set is classified by the Random Forest algorithm. However, the computational complexity as well as the increased dimensionality and redundancy of data sets based on morphological profiles are potential drawbacks.In the presented study, the method is used for the classification of urban hyperspectral data. The spatial information is derived by applying a series of opening and closing transformations to be the principal components of the original data set. However regarding the high redundancy and perhaps irrelevant information, the use of feature extraction methods seem appropriate. Here, we will investigate the use of two feature extraction methods, the nonparametric weighted feature extraction and the variable importance included in the Random Forests. For the classification, a freely available RF code was used in a MATLAB environment. The classifier training was performed with 50 training samples for each class and 100 trees within the RF. The number of variables used at each split was set to the square root of the number of bands.The MPs were created based on the first three principal components of the hyperspectral data set. The MPs used either a disk or a line as their SE, the radius/length ranged from 2, 3, 4, and 5. In addition the direction of the line was varied between 0, 45, 90, 135 degrees.FE was applied on both the MPs and the original hyperspectral data. Afterwards several classifications were performed, using the original data set, the MPs and data sets with reduced dimensions. Overall, five different classifications were performed: 1) Original image 2) Original image + morphology 3) Original image + morphology using feature selection a) RF variable importance b) NWFE c) NWFE + RF variable importanceThe use of feature extraction is seen to improve the classification accuracy, compared to the results achieved with the original data set and the spectral-spatial data set without any features selection. Whereas an accuracy of 66.2% is achieved with the original data, the accuracy increased up to 73.7% using a data set with a reduced dimensionality, underlining that feature reduction is important in the given context.Feature extraction can help in finding the most important features in both the original data set and the morphological profiles and thus, reduces redundancy in the data set.However, using the NWFE on both, the original data set and the MPs the improvement in the accuracy is less, as compared to the approach which is based on both feature extraction methods. Thus, it seems more adequate to apply the NWFE on the original data sets and use the VI for the feature selection on the MPs. The highest overall accuracy (73.7%) was achieved by selecting the 10 most important features of the original data set, using NWFE, and selecting 60 features out of the morphological profiles by RF's Variable Importance.Thus, in the presented study, feature selection is applied, using nonparametric weighted feature extraction and the variable importance of the random forests. The proposed approach is applied to ROSIS data from an urban area. The experimental results demonstrate that a feature reduction is useful in terms of accuracy. Moreover, the proposed approach also shows excellent results with a limited training set.Spectral-spatial classification of hyperspectral data has been investigated. Experimental results of the proposed spectral-spatial approaches on a ROSIS data set gave very good accuracies and excellent improvements compared to those obtained with pixel-based classifiers. Moreover the results clearly demonstrate the advantage of feature extraction in terms of accuracy. Based on the results we can also conclude that the use of the NWFE is helpful for the classification in the original image, but the Random Forest variable importance is more adequate for the feature selection of the morphological profiles.