Image Classification 

(Unsupervised  and Supervised classification)

by :  Idung Risdiyanto

       MIT Student / Biotrop

       1999

Brief Explanation

Background and definition

            Reflected electromagnetic radiation each object on the surface was received by satellite sensor has characteristic. Therefore, if we sure about image enhancement, we can make classification of the object on the surface was we want. For example, if we want to analysis object of agriculture planning only, we can classify the object on the surface was have relationship with agriculture like as field, river and forest.  We do not need other object if it object do not have relationship with agriculture.  So that, classification of land cover will help us to analysis without to make confusing with undesired object. According Jenson, 1986, it is possible to analyze remote sensed data of the earth and extract useful thematic information.  Notice that data are transformed in to information.  One of the most often use method of information extraction is multispectral classification. And then in Lillesand et al (1994) talk about the overall objectives of image classification procedures is to automatically categories all pixel in an image in to land cover classes or themes. Generally, the objectives of classification operations are to replace visual analysis of the image data with quantitative technique for automating the identification of features in a scene.  This normally involves the analysis of multispectral image data and the applications of statistically based decision rules for determining the land cover identify of each pixel in an image.  Before we make the image classification, we should be refers to several definition about them.  In these below are several definitions about image classification of land cover (O’Brien, 1999).

Image space - refers to the spatial/geometric locations of features

feature space - multidimensional space with as many dimensions as spectral bands measured

spectral pattern recognition - categorization of pixels based on their spectral information and the location(s) of features in multidimensional feature space.

Spatial pattern recognition - categorization of pixels based on their “visual similarities” and spatial relation to surrounding pixels.  Such similarities include texture, color, size, shape, proximity, repetition, direction, etc.  These similarities can usually be discerned by a manual interpreter.

Temporal pattern recognition - categorization of pixels based on understood and distinct spectral and spatial changes over time.   This includes known behavior over growing seasons, tidal and water level fluctuations, etc.  This technique is usually used to assist classification where a distinction may not be possible with an image from a single date.

Unsupervised classification - classification based on the aggregation of spectral clusters, followed by a manual interpretation of the cluster identity.

Clusters - natural spectral grouping in feature space

Supervised classification - classification based on the interpretation of an image to establish a numerical description of a class with training areas, obtaining the spectral characteristics or signatures of each class, and the automatic classification using the spectral signature information

training areas - representative sample sites of known cover types

omission errors - error representing pixels that were wrongly omitted from a category

commission errors - error representing pixels that were wrongly included in a category

Type of image classification

            In this assignment, we will make classification using three main type classifications. Each of classification type has advantages and disadvantages. First, two-band features space classification. This classification use histogram examination, slicing and scatter diagram exploration technique to perform simple classification.  If we use this type, we may be loss pixel or not all pixel in desired object can be classified, but this type will be helpfully if we make classify a few objects.  This type may be has big probability of overlap features.  Second type is unsupervised classification.  This type often use if we do not know or not familiar about our image or location.  The computer can be automatically to classify pixel base on different of DN value and statistic of the image. This type requires only a minimal amount of initial input from the analyst.  It is process whereby numerical operations are performed that search for “natural” groupings of the spectral properties of pixel, as examined in multispectral features space (Jenson, ibid). Then, the last one is supervised classification.  If we can be know or familiar with location or other object specific on the image we can use this type to classify. The object or locations were we know we called the training area. 

According in Jenson (ibid), the following are important aspect of conducting a rigorous and hopefully useful supervised classification of remote sensor data:

Supervised classification has three mains the appropriate classification algorithm – The parallelepiped, the minimum distance to means and the maximum likelihood classification.  Each of algorithm have a advantages and disadvantages. Therefore, in this assignment we will compare these algorithms, especially minimum distance to means and maximum likelihood, in order to know and understand what the different including advantages and disadvantages them.

These figure are few example of Classification using scatergram

Band 1 and 4 combination (Low correlation)

Band 3 and 5 combination (Moderate correlation)

Band 3 and 7 combination (high correlation)

References

Jenson, J.  1986.  Introduction Digital Image Procesing. A Remote Sensing Prespective.  New Jersey : Prentice - Hall

Lillesand, T.M and Ralph, W.K.  1994.  Remote Sensing and Image Interpretation (third edition).  John Wiley & Sons, Inc. New York. USA