When using unsupervised classification, pixels are automatically classified into a user-specified number of image classes according to their spectral properties, after which the classes are manually labelled (Campbell 2007). The exploratory nature of this automated spectral delineation, allows repeated unsupervised area delineations with different parameters, enabling users to “get a feel” of which real-world classes are spectrally distinct and similar (Mather 2004). Unsupervised
classification is extremely useful where a priori information regarding the study area or the classification structure is unavailable or not pre-determined (Campbell 2007).
On the other hand, supervised classification is defined by the application of a priori information of informational classes to determine the identity of unknown image elements. Data for informational classes are supplied to the classifier in the form of “training data”. These externally sourced training data is used to obtain statistical information regarding the spectral properties of each class, which is then used by a classification algorithm to identify the class of unknown pixels (Campbell 2007; Mather 2004). Various classification algorithms are available, with the most used algorithm being the maximum likelihood classifier (MLC) (Tseng et al. 2008; Brown de Colstoun et al. 2003; Bolstad & Lillisand 1991). The classifier statistically compares the features of each of the known classes with those of an unknown pixel in geometric space, and assigns the pixel to a class based on the results of the comparison. Other classifiers such as artificial neural networks and support vector machines are gaining popularity. An overview of each of these classifiers can be found in Pauw (2012).
While supervised classification has some advantages and produce acceptable accuracies, supervised classification is only as accurate as the training data used. Training data must be therefore carefully prepared, which can be costly (Albert 2002). In addition, traditional supervised classifiers are often out-performed by more elaborate classification methods, such as artificial neural networks, expert systems and DT (Pal & Mather 2003).
A DT classifier recursively applies a set of decision rules to an input dataset, categorising the dataset into a set of target classes. A decision tree classifier is composed of a root node (the input dataset), internal nodes (splits) and terminal nodes (the target classes, known as leaves). Although each node in the tree can only have one parent node, there is the possibility of having two or more descendant nodes. Decision rules are applied at each non-terminal node, splitting the data into smaller subsets until the leaf nodes are reached and the data were classified (Chuvieco & Huete 2010; Friedl & Brodley 1997).
Various approaches have been proposed for the construction of a DT. Rules can sometimes be created manually based on the analyst’s experience. According to Chuvieco & Huete (2010), such approaches can also be regarded as expert systems. However, supervised approaches where statistical procedures are used to infer the rules from training data, known as learning algorithms (Chuvieco & Huete 2010; Friedl & Brodley 1997), are commonly used.
Compared to other classifiers, DTs have the advantage of being able to accept a wide variety of input data, including both continuous and categorical data. Thus, ancillary data can easily be included. The simplicity of the structure of the classifier enables easy interpretation, straightforward testing as well as refinement when needed (Brown de Colstoun et al. 2003). Rogan, Franklin & Rogers (2002) found a DT to be significantly more accurate than a MLC in monitoring changes in forest vegetation in California. Similarly, Friedl & Brodley (1997) reported that DT provide significantly higher accuracies than MLC. Brown de Colstoun et al. (2003) also noted a higher accuracy for a decision tree classifier than that of traditional classifiers and comparable to that of a neural network (ANN).
The term “expert system” is used in many ways in remote sensing and can represent a number of different techniques. Tsatsoulis (1993) defines the categories of expert systems as user-assistance systems, classifiers, low-level processing systems, data fusion systems, and GIS applications. All pertain to different procedures in remote sensing analysis, but all are defined as “expert” in that they all employ artificial intelligence (AI) inference structures which use expert knowledge (Cohen & Shosheny 2002). For this reason, expert systems are also known in the literature as knowledge- based systems.
In contrast with traditional pixel-based classification approaches, object-based image analysis (OBIA) aims to delineate meaningful spatial units and classify them in an integrated way (Lang 2008), thereby making use of spatial concepts (Blaschke et al. 2000). In OBIA, each image object is aware of its context, neighbourhood and sub-objects so that geographical features can be characterised by their spatial, structural and hierarchical properties in addition to their spectral properties (Lang 2008; Bock et al. 2005). Furthermore, since objects offer additional spectral information including mean, median, minimum, maximum and variance values (Blaschke 2010), using objects as classification units rather than pixels reduces spectral variation within classes and removes the so-called “salt-and-pepper” effect (Liu & Xia 2011). In addition, OBIA is an iterative process which links to concepts of multi-scale analysis (Lang 2008; Blaschke 2010).
Several classifiers have been successfully applied to object-based classifications. Rule-based expert systems feature prominently in the literature on OBIA (Lang 2008). Expert systems attempt to model the complex network of knowledge and experience that humans use to understand the information in an image based on our perception of an image as a series of objects (Blaschke et al. 2000). Several studies report good results when using expert systems in object-based classification (Chen et al. 2009; Whiteside & Ahmad 2005).
For sparse, heterogeneous and largely non-photosynthetic material such as cleared land or newly restored patches, Steele et al. (2013) suggest using OBIA with decision-tree classification to provide a platform for classification that closely resembles human recognition of objects within a remotely sensed image, using the spatial context of the pixel and its relationship with its neighbours. However, Forsyth (2012) found that for single invasive species within a mountain landscape, pixel-based supervised classification performed better when compared to OBIA. Segmentation is the initial step in object-based image classification where multiple pixels in an image are consolidated and segregated as discrete units (Kartikeyan, Sarkar & Majumder 1998). Segmentation algorithms basically amalgamate adjacent pixels iteratively into larger objects at certain homogeneity threshold defined by the user. Mostly selected according to the demands of a particular task, segmentation algorithms generally improve accuracy of classification process and can be optimised by varying the selected region, adjusting threshold of both merged regions and merging termination (Blaschke 2010; Walter 2004). They are often categorized as either unsupervised or supervised. While unsupervised approaches use spectral data to identify areas with common attributes which are then assigned to arbitrary land cover classes, supervised classification methods use expert system to place pixels or areas into predefined land cover classes (Lillesand, Kiefer & Chipman 2004).
Segmentation pitfalls such as of over- and under-segmentation have the potential to falsify the classification process and this occurs mostly as a result of radiometric noise and inelasticity of homogeneity variables (Baatz, Hofmann & Willhauck 2008). While under-segmentation incorporates surplus surroundings not in comparison with the selected segments into the output, over-segmentation erroneously omits certain variables from the output (Stuckenberg, Münch & Van Nierkerk 2013). However, these errors typically occur where multiple classes are consolidated into a unit feature in a segment; where features of a land cover are embedded in a more dominant land cover class; and where boundaries of segments are not corresponding to features on the ground. Although several techniques are available to diffuse the effects of these segmentation pitfalls, the occurrence of these aberrations are likely to manifest to varying degrees in any segmentation exercises (Stuckenberg, Münch & Van Nierkerk 2013).