Object-Based Classification of Wetlands Using Optical and SAR Data with a Compound Kernel in Support Vector Machine (SVM)

Abstract:

Both Synthetic Aperture RADAR (SAR) and optical imagery play a pivotal role in many applications. Thus it is desirable to fuse the two independent sources of data congruously. Many of the fusion methods, however, fail to consider the different nature of SAR and optical data. Moreover, it is not straightforward to adjust the contribution of the two data sources with respect to the application. Support Vector Machine (SVM) is one of the classification methods which can provide the possibility of combination of two kinds of images considering the different nature of them. It is particularly useful when object-based classification is used, in which case features extracted from SAR and optical images can be treated differently.

This paper aims to develop an object-based classification method using both optical and SAR data which treats the two data sources independently. For the implementation of the method, a RapidEye and a RADARSAT-2 Quad-polarimetric image over Avalon Peninsula in Newfoundland, Canada will be used for wetland classification. RapidEye will be segmented using multiresolution algorithm in eCognition^TM. Because of speckle, segmentation of SAR images does not have robust results. Thus the result of the segmentation from RapidEye image is superimposed on RADARSAT-2 image. Then useful SAR and optical features are extracted. Integrating features extracted from optical and SAR data, a compound kernel in SVM is applied for classification. This kernel is a combination of two kernels with different weights, each of which are for the features of one of the data sources. Using compound kernel can outperform using the same kernel for both images.

The proposed method has two main advantages. First, different nature of optical and SAR images which is the result of dissimilar dynamic range, resolution, etc. is considered. Second, as the two data sources are combined with different weights, it is possible to adjust the role of each data sources for varying applications.