10.5 Object based classification

Objectives

The objective of this learning object is to introduce the principles of object based classification or Object based image analysis.

In the classification approaches discussed so far,  the unit at which the classification is applied is at the pixel level. However, when we have high spatial resolution images we can exploit this by grouping pixels in a manner that represents the way we view landscape features (i.e. as objects). Therefore, in object based classification, the unit of classification is the object.

In object based classification we try to group pixels that are similar based on measure of not only the spectral properties(i.e. colour/reflectance) but also size, shape and texture as well as context from the neighborhood (spatial autocorrelation).

Steps in Object classification

  1. Segmentation

The segmentation process splits an image into unclassified ‘object primitives’ that form the basis for the image objects and the rest of the image analysis. During the segmentation process, the size, colour, shape and pixel topology set by the user is taken into account. The values of these parameters determine the degree of influence of the spectral or spatial characteristics on the output objects.  The user modifies these settings depending on the objective of their task as well as the quality of the image available, the number of bands available and the image spatial resolution. As a general rule, good objects should be as large as possible, but small enough to show contours of interest and to serve as building blocks for objects of interest that are not yet identified. The ‘best’ settings for segmentation parameters can vary widely and often depend on a combination of trial and error and the expertise of the user.

When performing segmentation using most object based classification software(e.g. Definiens eCognition) you can choose the setting of various parameters such as colour, shape and scale.

The colour and shape parameters affect how objects are created during segmentation. The higher the value for colour the more the object will be optimized for spectral homogeneity whereas the higher the shape value the more the object will be optimized for spatial homogeneity. The colour and shape parameters balance each other, that is, if you have assigned a higher value for colour, then you need to assign a lower value for the shape and vice versa. If the two values are equal, then they will have roughly the same amount of influence on the segmentation outcome.

The scale parameter is an abstract value used to determine the maximum possible change of heterogeneity of objects. It affects the size of the objects created. High scale value leads to high variability within the segment and the objects tend to be relatively large. Conversely, small scale values allow less variability within each segment, creating relatively smaller segments. Homogenous areas often result in large objects, while heterogeneous areas often result in smaller objects.

A  number of segmentation algorithms have been implemented in several object based classification software. The two most commonly used segmentation algorithms include: region growing/merging and boundary detection algorithms . Region growing algorithms examine the neighbouring pixels of initial seed points and determines whether the pixel neighbours should be added to the region. Boundary detection algorithms look at abrupt changes in pixel values when grouping pixels together.

2. OBIA Classification

After segmentation, the objects are assigned to classes based on features (threshold to some characteristics e.g. colour, size, shape, texture) and criteria set by the user. The features usually define the upper and lower limits of a range of measures in the characteristics of image objects, e.g. how long is the shape, what are the levels of homogeneity in the texture etc. The image objects within the defined limits are assigned to a specific class and those outside these features are assigned to another class or are left unclassified. There are several methods of classifying the image after segmentation including: the Nearest Neighbour (NN) and Membership Function (MF) Methods.

  • The Nearest Neighbour (NN) Method

In the NN method, the user chooses a sample of  image objects that represent the classes they want to produce. The samples are based on prior knowledge of the study site features and should represent a range of characteristics within a single class. These sample objects are then compared to the image using some statistical measure e.g. comparing the mean shape, mean compactness, length of the shape, etc. and images are classified based on the closeness of these features to the sample objects. This is done repetitively until the whole image is classified.

  • Membership function (MF)

The second  most commonly used approach in object based classification is the membership function approach. Here, the user chooses different thresholds of various features (e.g. shape, texture) which must be satisfied before an object is classified into that class. The software then separates the image objects into classes using these feature thresholds. The membership function approach implies that one object may belong to more than one class as its qualification for a class is based on some level of percentage membership of that class. This approach is useful if classes can be separated by one or more features. It is appropriate when little prior knowledge is available.

Several software  packages are available to implement object based classification including: Definiens eCognition, ENVI and ArcGIS.

Try out object based classification/feature extraction using ENVI by following this practial here

 References

Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS journal of photogrammetry and remote sensing, 65(1), pp.2-16.

Walter, V., 2004. Object-based classification of remote sensing data for change detection. ISPRS Journal of photogrammetry and remote sensing, 58(3-4), pp.225-238.

Liu, D. and Xia, F., 2010. Assessing object-based classification: advantages and limitations. Remote Sensing Letters, 1(4), pp.187-194.

Comments are closed.