Читать книгу Change Detection and Image Time-Series Analysis 1 - Группа авторов - Страница 14

Depending on the purpose of unsupervised CD tasks, two main categories of methods are defined: binary change detection and multiclass change detection. The former aims to separate only the change and no-change classes, whereas the latter detects changes and distinguishes different classes within the changed pixels. In this chapter, we consider the latter, which is more attractive but challenging in practical CD applications. Note that in the unsupervised CD case, no ground truth or prior knowledge is available, thus the data-driven CD process is more preferable than the model-driven process. Therefore, the multiclass discrimination represents the inter-change difference associated with specific land-cover class transitions, whereas the detailed “from–to” information is absent, making it essentially different from the supervised case.

In general, the unsupervised CD process includes the following main steps: (1) multitemporal data pre-processing; (2) feature generation and selection; (3) change index construction; (4) CD algorithm design; (5) performance evaluation. The main components of an unsupervised CD are shown in Figure 1.1. Each step is briefly described and discussed as follows.

Figure 1.1. The main technical components of an unsupervised CD process

Multitemporal data pre-processing: in this step, different operations such as calibration, band stripe repair (if any), radiometric and atmospheric corrections, image enhancement and image-to-image co-registration are usually conducted in order to generate high-quality pre-processed multitemporal images for CD in the next steps. In particular, a high precision of co-registration is the core operation for a successful CD, which may significantly affect the CD performance due to the presence of remaining residual errors.

Feature generation and selection: features extracted from original multitemporal images are the critical carrier for representing different characteristics of objects in the single-time image and their variations in the temporal domain. Features such as original spectral bands, spectral indices (e.g. Normalized Difference Vegetation Index – NDVI, Modified Normalized Difference Water Index – MNDWI, Index-based Built-up Index – IBI) and textures (e.g. mean, contrast, homogeneity) derived from original bands can be considered in CD. In addition, spatial features generated from multispectral bands such as wavelet transformation (Celik and Ma 2011), Gabor filtering (Li et al. 2015), morphological filtering (Falco et al. 2013), etc., provide important multi-scale geometric information about image objects to improve the change representation. Recently, deep learning-based CD approaches have shown great potential in extracting more high-level deep features, which represents a popular direction in CD research (Mou et al. 2019; Saha et al. 2019).

Change index construction: the change index represents the temporal variations extracted from multitemporal image features. It can be constructed based on different operators and algorithms, such as univariate image differencing (Bruzzone and Prieto 2000a), change vector analysis (CVA) (Bovolo and Bruzzone 2007b), ratioing (Bazi et al. 2005), distance or similarity measures (Du et al. 2012), etc. Transformation approaches such as iterative reweighted multivariate alteration detection (IR-MAD) (Nielsen 2007), principal components analysis (PCA) and its kernel version (Nielsen and Canty 2008; Celik 2009), independent component analysis (ICA) (Liu et al. 2012), are also designed to transform the change information from the original data space into a projected feature space. However, a careful selection of specific components representing user-interested changes is required. This is often very difficult in an unsupervised CD case without prior knowledge about the considered study area and dataset, which may limit the automation degree of the CD application. For a summary of the related methods for constructing different types of change index, readers can refer to the paper by Bovolo and Bruzzone (2015).

Change detection algorithm design: unlike the supervised and semi-supervised CD methods that rely on the available reference samples, unsupervised CD algorithms focus more on the automation and accuracy. Thus, basically, most of the unsupervised CD approaches are data-driven by analyzing the multitemporal data itself. Within this context, for binary CD, if we consider a given change index generated in the previous step, for example, the magnitude of differencing image, automatic thresholding such as empirical segmentation (Bruzzone and Prieto 2000b), Kittler–Illingworth (KI) (Bazi et al. 2005), Otsu (1979) and Bayesian-based expectation–maximization (EM) (Bruzzone and Prieto 2000a) are all simple but effective algorithms proposed in the literature. However, the successful use of such methods depends on the assumption of a certain data distribution such as Gaussian or Rayleigh–Rice mixture (Bovolo et al. 2012; Zanetti et al. 2015), where a wrong estimation may lead to many detection errors. On the contrary, clustering algorithms such as k-means, fuzzy c-means and Gustafson–Kessel clustering (GKC) have been used to address the binary CD problem (Celik 2009; Ghosh et al. 2011), which are distribution-free but require a specific setting to avoid unstable performance, such as the accuracy decrease due to random initialization.

For a multiclass CD case, the unsupervised task becomes more complex since several sub-problems should be solved simultaneously, including the binary change and no-change separation, the number of multiclass change estimation and the multiclass change discrimination (Liu et al. 2019c). In particular, among many solutions, we recall the classical multiple CD technique – change vector analysis (CVA) (Malila 1980). It was designed to analyze possible multiple changes in pairs of bitemporal image bands. A theoretical definition was given to the original CVA approach in the polar domain to provide a more clear mathematical explanation to CVA (Bovolo and Bruzzone 2007b). However, it still has a limitation, i.e. only a part of all possible changes can be detected since only two selected bands are considered in each implementation. If more spectral channels are considered, it becomes very difficult to simultaneously model and visualize multidimensional changes. To break this constraint, a compressed change vector analysis (C²VA) approach was proposed, which successfully extended the original CVA to a two-dimensional (2D) representation of the multi-band problem (Bovolo et al. 2012). Other works in the literature developed different variations of CVA. For example, a modified CVA was developed to determine the magnitude threshold and direction by combining single-date image classification results (Chen et al. 2003). An improved thresholding approach on change magnitude was designed to optimize the binary separation on each specific change class (Bovolo and Bruzzone 2011). A hierarchical version of C²VA with an adaptive and sequential projection of spectral change vectors (SCVs) at each level of the hierarchy was proposed to detect multiple changes in bitemporal hyperspectral images (Liu et al. 2015). In this chapter, we also explore the potential capability of C²VA and extend it from the spectral–spatial point of view.

Performance evaluation: similar to the supervised CD methods, unsupervised binary and multiclass CD approaches can usually be assessed according to the detection accuracy or error index, such as overall accuracy (OA), Kappa coefficient (Kappa), omission errors (OE) and commission errors (CE), receiver operating characteristic (ROC) curve and area under the curve (AUC) value. In this case, the accuracy evaluation usually relies on the manually interpreted change reference map. Note that such a reference map is only used for accuracy evaluation, which is not considered as training data as in the supervised case. In addition, the computational time cost is also another important indicator that reflects the automation and efficiency of unsupervised methods.