10.2. Image Classification Types

---

Objectives

The objective of this learning object is to introduce the principal approaches to image classification including unsupervised, supervised and object based classification.

A key distinction between image classification approaches is whether or not the analyst uses prior information or whether the classification is based entirely on the data themselves. In the latter case, decisions are still required to be made, but these relate to the process and not to the details of the classification itself – this is termed unsupervised classification. By contrast, if the user already has some information about the land cover types at different locations in the area represented by the image, these can be used to ‘train’ the classification process – this is termed supervised classification.

We can here summarise some of the principal advantages and disadvantages of these approaches:

Unsupervised Image Classification (UC)
Advantages (relative to supervised classification)	Disadvantages (relative to supervised classification)
No extensive/detailed a priori knowledge of the region is required	Spectral grouping produced by the classifier may not correspond to the information classes of interest to the analyst
Human error and bias is minimised (fewer decisions are required by the analyst)	There is limited control over the ‘menu’ of classes
More uniform classes are produced	Spectral properties of specific classes will change over time (relationships between information classes and spectral classes are not constant)
Spectrally distinct classes present in the data may be revealed which were not initially apparent to the analyst

Supervised Image Classification (SC)
Advantages (relative to unsupervised classification)	Disadvantages (relative to unsupervised classification)
The analyst has full control of the process	Signatures are forced, because training classes are based on field identification and not on spectral properties
Processing is tied to specific areas of known identity	Training data selected by the analyst may not be representative (heterogeneity within classes is common)
The analyst is not faced with the problem of matching categories on the final map with field information	The preparation of training data is time-consuming and costly (an iterative process)
The operator can detect errors, and is often able to remedy them	It is not possible to recognise and represent special or unique categories which may be present in the image but are not represented in the training data

---

Reflection