Utilization of Spectral Information in Image Categorization
The utilization of quantitative spectral information present in images is used in image categorization. The information should be connected to the target surface's conditions or compositions. Image analysis can be performed on both hyperspectral and multispectral imagery. Understanding how objects and materials of interest on the earth's surface reflect, emit, and absorb radiation in the near-infrared, visible, and thermal sections of the electromagnetic spectrum is required for picture categorization. To perform correct picture analysis in a GIS setting, the source image should be orthorectified before being layered with other topography data, images, and geographic data layers. The classification results are usually in raster format but they can also be generalized to form polygons in further analysis. There are numerous essential principles of image analysis that relate specifically to the extraction of features and information from remote sensing data. The remote sensing data normally provide critical and essential information for analyzing different applications like image change detection, land cover classifications, and image fusion. Remote sensing is an imperative technique for obtaining information that relates to the Earth's environment and resources. The popularity of the satellite data came about as a result of easily accessible online mapping applications like Bing Maps and Google Earth. These applications have enabled the researchers to locate different areas within the earth and they have also enabled the GIS community in monitoring disasters, program planning, natural calamities, guiding civil defense people, and project planning. The two major image classification techniques include the supervised image classification technique and the unsupervised image classification technique; the paper, therefore, discusses the strengths and weaknesses of the above classification techniques.
Unsupervised Classification
The unsupervised classification of data involves the use of pixels. The pixels are normally grouped according to the reflectance properties. The grouping is usually referred to as clusters. In the unsupervised classification, the user normally identifies the total number of clusters to generate as well as the bands to use. With the above information, the image classification software usually formulates structures. There are various image clustering techniques like the ISODATA and K-means. In the unsupervised classification, the user manually categorizes each cluster with land cover classes; it is usually the case that several clusters signify one land cover class. The user often merges clusters into various land cover types. The unsupervised image classification technique is frequently used when there is a lack of sample sites. In the unsupervised classification technique, the users manually identify each of the available structures together with land cover classes. Multiple clusters embody a single land cover class. The users normally fuse clusters into a land cover type.
Strengths and Weaknesses of Unsupervised Classification Technique
Strengths
Unsupervised classification is essential in evaluating areas that are unpopular. In most cases, the users may lack appropriate knowledge about the anticipated sites; this, therefore, makes it necessary to apply the unsupervised classification technique. The above technique can also be used as the initial tool for assessing the scene before the actual application of the supervised classification system. The use of the unclassified classification system is usually unbiased when it comes to the geographical assessment of the pixels. Consequently, the supervised classification system enables the users to manually identify the training sites before the actual operation begins. The advantage of both the unsupervised and the supervised classification techniques lies with the efficiency by which the programs can be used to undertake the statistical analysis. Once the pixel clusters have been assigned, it is easier to identify the specific number of pixels in each of the classes. As the size of each of the pixels from the available metadata, the metric position of each of the classes can be easily calculated.
Weaknesses
In most cases, the use of unsupervised classification techniques is usually done without the knowledge of the site; this can make the formulation of the algorithm to be so difficult. In some other cases, the logarithms may become irrelevant during the actual application in different sites. For example, lack of knowledge of the scene may cause the user to carry out various experimental activities to identify each of the spectral structures. Each interaction may become time-consuming, and also, the final image may become difficult to interpret specifically when there is a large number of unidentified pixels. The unspecified classification is insensitive to variations and covariations in the spectral signature to objects. The algorithm may erroneously differentiate pixels with slightly different spectral values and assign them to different clusters when they may be representing a spectral feature of a group of the same objects.
It is common for a study to include areas beyond the range of a personal scene. In the above cases, it is essential to collect mosaic or adjacent scene and place them together. It is normally preferable to identify scenes with the collected data obtained during the same time frame or season and under similar weather conditions. Images can be pieced together of the data collected are validated in the same datum and projection. It is also necessary to assess the registration of all the available images before filing the scenes together. The misregistered images may lead to gaps in the image clusters or may lead to pixel overlay.
Supervised Classification Technique
In the supervised classification technique, the user normally selects statistical samples for each of the land cover clusters inform of the digital image. The sample land cover clusters are known as the training sites. The classification of the image requires an application that uses the training sites to locate the land cover classes in the whole of the image. The classification of the land cover is often based on the spectral signature formulated in the training sites. The image classification software identifies each class on what is resembles mostly in the training field. The commonly supervised classification technique algorithms are minimum-distance and maximum-likelihood classifications. The traditional pixel-based processing creates square forms of pixels. The object-based image classifications are very diverse because they generate objects of different scales and shapes. The above process is called multi-resolution segmentation. The multi-resolution differentiation normally produces homogenous image objects by categorizing pixels. Objects are commonly generated by diverse scales in an image simultaneously. The above objects are often meaningful than the traditional pixel-based division because their classifications can be done based on the texture, geometry, and context.
Strengths and Weaknesses of Supervised Classification Techniques
Strengths
The supervised classification involves the pixel classification processes; the image analyst or the user supervises the classification processes. During the classification process, the user specifies values of various spectral signatures and pixels that should be linked to each class. The above process is done by identifying the representative sample sites of the known cover type usually called areas or training sites. The computer application then uses the spectral signatures from the identified training areas in order to provide an accurate image. Preferably, the classes should not overlap or should overlap minimally with other classes.
In the supervised classification technique, a lot of effort is done before the actual classification is done. After the classification is done, the output is produced in the form of a map with clusters that are labeled and correspond to the information clusters and land cover types. The supervised classification is often much more accurate when compared to the unsupervised classification, although it sometimes depends heavily on the training sites, individual skills in processing the images, and the spectral distinctness of the clusters. If more than one class is the same in terms of the spectral reflectance, the misclassifications may tend to be high.
Weaknesses
The supervised classification needs close attention to the training of knowledge on the developmental data. If there is poor data or poor representation of data, the classification may also be poor. Therefore, the supervised classification mostly requires more money and time when compared to the unsupervised classification. Multi-resolution segmentation results in homogeneous image objects by categorizing pixels. These images are more complex and detailed when compared to the traditional classification systems. The above scenario is due to the complexity of the processes involved in the analysis processes which normally results in the more complicated images. The supervised classification is commonly related to the Nearest neighbor classification where the user identifies sample sites for each cover class; this usually makes the classification more complicated as more time and more are required in the analysis process. The processes of producing the accurate images of the expected sites usually require specialized equipment which may be expensive and difficult to find in some areas. On the other hand, the methods of producing data often depend on the accuracy of the training data; this means that when the data is not accurate, the resulting images will have low qualities. The production of the data also follows the representation of the training data, a situation that requires double checking and accuracy. Lastly, the accuracy of the collected data depends on the distinctiveness of the classes which need to be observed right from the beginning of the research.
