Classification of high resolution pixels into non-heterogeneous groups using rule based fuzzy approach

In remote sensing applications, the classification of pixels into non-heterogeneous groups is an essential task, where each group belongs to some particular land cover type. This problem is modeled as a supervised classification or a clustering problem. However, it requires having exact number of the true clusters in satellite images. The cluster algorithms require to be able to find the number of clusters. But, in satellite images, some clusters occupy very few pixels, while some neighboring regions or groups of such, widely varying sizes, presents a challenge in designing segmentation algorithm. Many supervised classification techniques exist in the literature to solve the image classification problem. It is also possible to classify the remote sensing image to clearly describe certain land feature classes such as roads, streets and other features. As a prerequisite, certain land feature classes with spectral information with known values is to be used in the fuzzy classification. An integrated method consisting of maximum likelihood and fuzzy classification is proposed in this paper. In this method, a suitable partitioning is detected in an iterated manner. In this paper, first maximum likelihood (ML) supervised and then fuzzy classification is done. The fuzzy rule sets are framed to identify the certain land features. Certain linguistic variables are used along with membership functions. The estimated area (% of pixels) of class is calculated, according to rules defined in the fuzzy classification system. The resultant classes of both the procedures are compared with k-means, isodata, ML classification algorithms. The accuracy of analysis is done by comparing percentage of pixels occurred in the classes. The fuzzy based classification method, found to give better results. This technique could be applied in several applications such as land use classification, hydrological applications, soil mapping, clustering, decision making, and operations research.