Polarimetric classification of Boreal forest using nonparametric feature selection and multiple classifiers

Abstract

Polarimetric SAR data contains a large amount of potential information that may be used to characterize forested scenes. However, the large number of PolSAR parameters and discriminators cannot all be used in most classification problems. Some form of feature selection will improve classification results and improve the efficiency of the system. In addition, classification of PolSAR data may be improved with an ensemble of classifiers, each tuned to a different class. Our research is in the Petawawa experimental forest, in the boreal forest northwest of Ottawa, Ontario, Canada. We employ Radarsat-2 fine-quad image data acquired in August (leaf-on) and November (leaf-off) of 2009. We present two system designs in this paper. The first system consists of a feature selector based on a non-parametric evaluation function and a support vector machine for classification. We demonstrate that the feature selection step improves classification accuracy significantly over a baseline classifier. We then present a system consisting of an ensemble of SVM classifiers, each with its own feature selection component and trained on an individual class. The classifier likelihoods are combined in a final step. We demonstrate that this system improves classification accuracy significantly over a single-classifier system. Finally, we demonstrate that classification accuracies are significantly higher when leaf-on and leaf-off images are combined over a single season image.

Introduction

Forests are a major natural resource of the Earth and control a wide range of environmental processes. Forest ecosystems play an important role in energy exchange between the land surface and the atmosphere (Jarvis and Dewar, 1993). They comprise a major part of the planet's plant biodiversity and have an important role in the global hydrological and biochemical cycles. In addition, forests are considered as an important sink of atmospheric carbon and an important component of any carbon cycle modeling. As well, demand for forest products is increasing worldwide particularly affecting Canada which has 10% of the world's forested area. Remote sensing data can be a valuable tool in forestry. There are numerous potential applications which include biophysical variable estimation, species discrimination or classification, disturbance detection, forest monitoring and ecological modeling (Kasischke et al., 1997). In this study we focus on species discrimination which is also called forest mapping.

Forest mapping is one of the core applications in remote sensing. Many studies are based on multispectral optical images, and the use of hyperspectral data are increasing due to their increased information content. Unfortunately, atmospheric conditions often limit the use of optical data. Synthetic Aperture Radar (SAR) are active systems that use longer microwave wavelengths, able to penetrate clouds and precipitation. These systems are largely independent of weather conditions and are also able to obtain images day or night. The longer wavelengths allow the use of polarized waves, since the intervening atmosphere does not change the polarization state of the transmitted or scattered waves. The majority of SAR systems use linearly polarized radiation and transmit and receive either horizontally (H) or vertically (V) polarized waves, and the data are denoted as XY, where X is the transmitted polarization and Y is the received polarization, e.g. HH means both the transmitted and received waves are horizontally polarized. The longer wavelengths and the addition of polarization information make radar observations sensitive to target geometry. These properties make SAR data a useful tool for many applications including forest mapping.

The extraction of information from SAR data has been an active area of research for since the launch of the first civilian SAR satellite, Seasat, in 1978. Despite certain challenges inherent in SAR imaging systems, such as their idiosyncratic geometry and speckle noise, SAR has become a standard tool for forest mapping. Early satellite SAR systems produced single channel co-polarized data (HH or VV), thus early investigations focused analyses these data. These single channel data were very useful often showing spatial features that were not visible in optical data. However, their use for classification studies were very limited. When images from different epochs were combined in a multi-temporal dataset it was possible to classify scenes based on the evolution of scattering properties over time (Quegan et al., 2000, Wegmuller et al., 2002). Other multi-channel studies using multi-frequency data (Kouskoulas et al., 1998), multi-polarization data (Du and Lee, 1996), and a fusion of SAR data and optical data (Bruzzone et al., 1999) have all demonstrated the power of combining radar channels.

More recently, SAR polarimeters have been used for remote sensing. These instruments generate quad-polarized data consisting of four linear polarizations, HH, HV, VH and VV, and the phase relationships between them. Spaceborne SAR polarimeters (PolSAR), such as Radarsat-2 and ALOS, have made quad-polarized SAR data available to the wider community leading to many general and specialized algorithms for extracting information from this rich dataset.

Classification of polarimetric SAR data has received considerable attention during the past decade, e.g. in (Cloude and Pottier, 1996, Cloude and Pottier, 1997, Lee et al., 1999, Lee et al., 2001) and also more recently in (Lönnqvist et al., 2010, McNairn et al., 2009, Lardeux et al., 2009). The richness of the polarimetric data has produced many approaches to analysis. Some use the observed data; either the scattering matrix, the covariance matrix or the coherency matrix. Another set of algorithms are based on some form of target decomposition (TD) theory. The main idea of these methods is to decompose the data into independent components representing different physical scattering mechanisms. Finally, there have been many descriptive parameters that have been designed to extract description information from quad-polarized SAR data. These three classes of algorithms form the basis of all PolSAR classification algorithms.

Classification studies focus on a single algorithm or class of algorithms. For example, the formalism developed by Cloude (1986) led to the introduction of a widely used unsupervised classification scheme (Cloude and Pottier, 1997), further augmented and improved by subsequent contributions (Pottier, 1998, Lee et al., 2004).

In spatially complex scenes, such as forests, it is useful to make full use of the discriminative power offered by the wide range of PolSAR features. However, due to the small sample size problem, using all these features in a classification is not feasible. Furthermore, some of these features likely carry redundant information. Therefore, a key stage in the design of a classifier system is the selection of the most discriminative and informative features. This paper demonstrates the advantages offered by adding a feature selection step in a classification system based on PolSAR features. We explore the use of feature selection in two systems, one with a single classifier and one consisting of a set of classifiers, one for each class.

There are a large variety of different algorithms that have been proposed in the literature for feature selection. For a recent overview see Pal and Foody (2010). For the case of our research, most of the PolSAR parameters have complex, and sometimes unknown, statistical properties. For this reason, the conventional parametric feature selection algorithms cannot be applied. To resolve this, a nonparametric feature selection (NFS) approach is employed for our single classifier system.

Upon the selection of the most appropriate features, they are transferred to the classification step. Recent work (Lardeux et al., 2009) reports the potential of the support vector machine (SVM) algorithm for the classification of multifrequency SAR polarimetric data. They showed that the SVM algorithm is particularly well suited to account for numerous and heterogeneous parameters, such as the intensity channels and different SAR polarimetric parameters. This, along with its ability to handle linearly non separable cases, a SVM algorithm is used in the classification step in our study.

We also apply feature selection to a multiple classifier system. In spite of the considerable amount of work that has been carried out on the use of an ensemble of classifiers for the classification in recent years, only a few applications have been reported for PolSAR data. She et al. (2007) applied Adaboost to PolSAR image classification. Compared with traditional methods, such as Wishart distance classifier, it was found to be more flexible and robust. Chen et al. (2008) proposed a supervised classification scheme based on Adaboost. Each independent element of the Mueller matrix as well as several extracted parameters formed a weak classifier. By the Adaboost procedure, each classifier was endowed with a weight. The weight presented the effectiveness of the corresponding feature. The weak classifiers were combined to form a strong classifier using a majority weighted voting. They found that this scheme was more robust and more accurate than the traditional maximum likelihood classifier. Min et al. (2009) also employed polarimetric decomposition and the Adaboost algorithm to solve a PolSAR image classification problem. Their simulation results validated their method compared with a standard H/α classification algorithm.

To create a classifier ensemble, we have developed a class-based feature selection (CBFS). In this framework, instead of selecting one set of features used for all classes, we select a separate feature subsets to discriminate each class. The SVM classifier is trained on these feature subsets and a combination scheme is used to combine the outputs of these classifiers.

The Wishart classifier, which is one of the most widely used methods for the classification of polarimetric data, as well as an SVM with full set of features were used as the baseline methods for comparison.

We also present experiments with the use of individual summer (leaf-on) and fall (leaf-off) scenes as well as the combination of leaf-on and leaf-off and show that the multi-temporal set yields the highest accuracy. Also, the usefulness of the selected PolSAR parameters for each dataset was examined in this framework.

The remainder of the paper is organized as follows: in Section 2 different polarimetric features are reviewed, in Section 3 the methodologies including the NFS, CBFS and the baseline methods are explained, in Section 4 the experimental data including the study area, reference data and the SAR data are described. Finally, the experimental results are presented and discussed in Section 5 and the conclusions are drawn in Section 6.