Abstract
Satellite imagery interpretation has become the technology of choice for a host of developmental, scientific, and administrative management work. The huge repository of geospatial data and information that are available as satellite imageries datasets from platforms such as Google Earth need to be classified and understood for natural resources management, urban planning, and sustainable development. The classification and analysis procedures involve algorithms like maximum likelihood classifier, isodata, fuzzy-logic classifier, and artificial neural network based classifier. Amongst these classifiers the optimum has to be selected for classifications which involve multiple features and classes. Herein lies the motivation for the present research, which can facilitate the selection of one amongst the many algorithms available to a decision maker/manager. The aforementioned techniques are applied for classification, and the respective accuracies in the classes of forestry, rock, water, built-up area, and dry river bed have been tabulated and verified from ground truth. The comparison is based on time and space complexity of the algorithms considering also the accuracy. It is found that traditional methods like MLC and Isodata offer good time and space consumption performance over the recent more adaptable algorithms as fuzzy and ANN. But the latter group excels in accuracy of assessment. The study suggests points and cases for ranking the techniques as best, 2nd best, and so on, where each technique could be optimally utilised for a given geospatial dataset based on its contents.
TopBackground
The algorithms considered for the study are all pixel based techniques that work on raster data sets in the domain of digital image processing and classification. Raster data is a two dimensional array of digital numbers (DN) where each cell of the array corresponds to a pixel of the image. The image could be a gray-scale (b/w) or a coloured image. Accordingly the array values would vary. A gray-scale would have a single layer of digital values and a colour image would consist of hues of RGB. Each DN value represents a unit intensity captured in the image (Lillesand & Kiefer, 1994).
The algorithms try to best utilize the categorization of these DN values. Maximum likelihood, isodata, fuzzy techniques and ANN have been used for numerous satellite image analysis studies but have not been compared amongst themselves on a single platform anywhere. The measures of comparison that effect the performance and utility most are space occupied in memory, time taken for processing, accuracy delivered, and future upgradability. To understand the study better, it is imperative to get a look into the backgrounds related to the study. In this section some common terminologies are discussed to lay the foundation for their applications later.
The simplest method of image classification is called the thresholding method. The key of this method is to select the threshold value (or values when multiple-levels are selected) for each category that exists in a digital image. Several popular methods are used in this field and in industry including the maximum entropy method, Otsu's method (maximum variance), and et al. Fuzzy “k-means” or simply “k-means” clustering can also be used. The present study analyses and compares MLC, Isodata, Fuzzy classification and ANN based classification.