Image Pre-processing Remotely Sensed Image Classification

9 The transformation of a remotely sensed image is called geometric correction or geo-referencing. A related technique, called registration, is the fitting of the coordinate system of one image to that of the second image of the same area. Accurate image registration is needed if a time sequence of images is used to detect changes in, for example, the land covers of an area date Mather 2004.

2.2.3 Image Processing

Image data, available in digital form, can be quantized spatially and radiometrically. There are several approaches are possible in extracting the information. Two approaches are usually used to extract information from digital image data: quantitative analysis or also called classification and photo- interpretation or sometimes called visual image interpretation. Photo- interpretation is aided substantially if a degree of digital image processing is applied to the image data beforehand, while quantitative analysis depends for its success on information provided at key stages by an analyst Richards and Jia 2006. Photo-interpretation which involves direct human interaction and therefore it needs high level decisions. It is good for spatial assessment but poor in quantitative accuracy. Area estimated by photo-interpretation, for instance, would involve planimetric measurement of regions identified visually; in which, boundary definition errors will prejudice area accuracy. By contrast, quantitative analysis, requiring little human interaction, has poor spatial ability but high quantitative accuracy. Its high accuracy comes from the ability of a computer, if required, to process every pixel in a given image and to take account of the full range of spectral, spatial and radiometric detail present. Its poor spatial properties come from the relative difficulty with which decisions about shape, size, orientation and texture can be solved by using standard sequential computing techniques Richards and Jia 2006. In computer-based quantitative analysis, the attributes of each pixel such as the spectral bands available are examined in order to give the pixel a label which identify it as belong to a particular class of pixels of interest to the user Richards and Jia 2006. The process of classification consists of two stages: 10 recognition of categories of real-world objects and labeling of the classified entities normally pixels. In the context of remote sensing of the land surface these categories could include, for example, woodlands, water bodies, grassland and other land cover types, depending on the geographical scale and nature of the study. In digital image classification the labels are numerical, so that group of pixels that are recognized as belonging to the class ‘water’ may be given the label ‘1’, ‘woodland’ may be labeled ‘2’, and so on Mather 2004. There are several methods used in image classification, but generally those methods can be categorized as unsupervised classification and supervised classification. Unsupervised classification is an analytical procedure based on clustering using some algorithms. Application of clustering partitions the image data in multispectral space into a number of spectral classes, and then labels all pixels of interest as belonging to one of those spectral classes. The process followed segmentation of the multispectral space to cluster pixels into ground cover types, by the analyst Richards and Jia 2006. Supervised classification methods are based on external knowledge of the area shown in the image. Unlike some of the unsupervised methods, supervised methods require some input from the user before the chosen algorithm is applied. This input maybe derived from fieldwork, air photo analysis, reports, or from the study of appropriate maps of the area of interest. In the main, supervised methods are implemented by using either statistical or neural algorithms. Statistical algorithms use parameters derived from sample data in the form of training classes, such as the minimum and maximum values on the features, or the mean values of the individual clusters, or the mean and variance–covariance matrices for each of the classes. Neural methods do not rely on statistical information derived from the sample data but are trained on the sample data directly. This is an important characteristic of neural methods of pattern recognition, for these methods make no assumptions concerning the frequency distribution of the data. In contrast, statistical methods such as the maximum likelihood procedure are based on the assumption that the frequency distribution for each class is multivariate normal in form. Thus, statistical methods are said to be parametric