Supervised change detection in VHR images using contextual information and support vector machines

Abstract

In this paper we study an effective solution to deal with supervised change detection in very high geometrical resolution (VHR) images. High within-class variance as well as low between-class variance that characterize this kind of imagery make the detection and classification of ground cover transitions a difficult task. In order to achieve high detection accuracy, we propose the inclusion of spatial and contextual information issued from local textural statistics and mathematical morphology. To perform change detection, two architectures, initially developed for medium resolution images, are adapted for VHR: Direct Multi-date Classification and Difference Image Analysis. To cope with the high intra-class variability, we adopted a nonlinear classifier: the Support Vector Machines (SVM). The proposed approaches are successfully evaluated on two series of pansharpened QuickBird images.

Introduction

One of the most challenging Earth observation task is the identification of land cover transitions and changes occurred on a given region. Land cover evolutions can be identified by the analysis of two or more coregistered remote sensing images of the same geographical area at different times (Singh, 1989, Coppin et al., 2004).

Nowadays, many commercial and governmental instruments provide images within small temporal intervals with high to very high spatial resolutions. This type of imagery is appropriate for the study and the analysis of localized ground cover changes. In the literature, several methods have been developed for this purpose and efforts were put in considering low and medium resolution imagery. In the last decade, many studies aimed at transferring this knowledge to high and very high geometrical resolution (VHR) images.

This paper focuses on VHR images and on the adaptation of existing automatic classification techniques to discover changes. Change detection is considered as a supervised multi-temporal classification problem, which aims at obtaining a complete description of the transitions occurred between the acquisitions. Moving to VHR imagery comes with the price of increased within-class variances, that prevent the successful application of traditional classification methods such as the Maximum Likelihood classifier. In VHR the use of a robust and nonlinear classifier is mandatory since noise and generally higher spread in class distributions makes the classification problem very complex.

Support Vector Machines (SVM) classifiers (Vapnik, 1998, Schölkopf and Smola, 2002, Shawe-Taylor and Cristianini, 2004) have demonstrated their effectiveness in several remote sensing applications (Camps-Valls and Bruzzone, 2009). In particular, several researches addressed the problem of VHR ground cover classification using SVM (Bruzzone and Carlin, 2006, Inglada, 2007, Tuia et al., 2009). The success of such approaches is related to the intrinsic properties of this classifier: can handle ill-posed problems and to the curse of dimensionality (Hughes, 1968), provides robust sparse solutions and delineates nonlinear decision boundaries between the classes.

Recently, kernel methods started to be considered also for change detection and multi-temporal classification. Despite the promising results in many remote sensing tasks, only few studies deal with change detection. In Nemmour and Chibani (2006) supervised multi-temporal classification is implemented using SVM. In their setting, two coregistered images are stacked and the bi-temporal dataset is classified with a multiple SVM approach. The comparison with a Neural Networks classifier proved that SVM are less prone to overfit the data and training issues related to non-convex error functions are avoided. Bovolo et al. (2008) perform transductive SVM for change detection initialized with a Bayesian selective thresholding method (Bruzzone and Fernández-Prieto, 2000) that allows the unsupervised application of this classifier. The final performance obtained outperformed classical change vector analysis. Bovolo et al. (2010) reformulated the change detection task as an outlier detection problem, modeling the target (changed patterns between the two times) via Support Vector Domain Description and detecting unchanged pixels as outliers. The superiority of the nonlinear approach was proven by their experiments.

As mentioned, in VHR images the underlying class distributions are often strongly overlapped, resulting in hardly classifiable pixels even using robust methods as SVM. The high within-class variance as well as the low between-class distance, due to the low spectral information, increase the need for approaches that enhance separability between the different classes. To solve this issue, contextual features providing information on the spatial relationships of pixels have been extensively studied for standard classification.

Spatial context features are often considered to ease the classification process of VHR images. Murray et al. (2010) proved that the joint use of spectral and textural features ameliorates the classification accuracy of VHR images considerably. On the opposite, classification performed using only spectral or textural features results in lower performance. In Tuia et al. (2009), different multi-scale morphological features are extracted and studied to classify QuickBird panchromatic images (thus with poor spectral resolution) using SVM. In Pacifici et al. (2009) local textural measures based on the Gray Level Co-Occurrence Matrix (GLCM) are studied for classifying VHR panchromatic images with a Neural Networks classifier. In Tuia et al. (2010b), specific kernel functions are designed to find optimal combinations of contextual information at relevant spatial scales. Summing up, these studies verify that the lack of spectral information is successfully balanced by the inclusion of contextual information.

The exploitation of spatial information is poorly documented in change detection literature, although the benefits of considering such variables are clearly demonstrated in classification tasks. In Dalla Mura et al. (2008) the advantages of including morphological reconstruction operators in the change vector analysis framework (Bovolo and Bruzzone, 2007) has been illustrated. By filtering the magnitude of the difference image (as an intermediate step), errors due to radiometric differences and noise are greatly reduced. In Bovolo (2009) a contextual parcel-based multi-scale approach to unsupervised change detection is presented. The usefulness of contextual information in VHR unsupervised change detection is clearly pointed out by these studies.

In this paper, we propose an effective way to deal with supervised change detection in VHR images by integrating spatial information in SVM multi-temporal classification. As introduced, it is already proven that the pixel context characteristics can provide accurate and coherent classification maps by filling the lack of spectral information. On the other hand, SVM are suitable tools for many remote sensing applications, thanks to their intrinsic properties. The rationale of this paper is to combine the advantages of both SVM and contextual information and to prove their benefits for supervised change detection in VHR images. This aims at mitigating class separability problems by completing the feature vector, and discovering the optimal nonlinear classification boundaries with SVM. Two change detection architectures are considered: Direct Multi-date Classification (DMC) and Difference Image Analysis (DIA).

The remainder of the paper is organized as follows: Section 2 introduces the reader to the extracted features, to the classifier and to the change detection architectures. Section 3 presents the datasets as well as the experimental setup. Section 4 presents results, Section 5 discusses the outcomes and Section 6 draws the conclusions of the paper.