Introduction |
Wetlands are an essential part of human civilization, meeting many crucial needs for life on earth such as drinking water, protein production, energy, fodder, biodiversity, flood storage, transport, recreation, and climate stabilizers. They also aid in improving water quality by filtering sediments and nutrients from surface water. Wetlands play a major role in removing dissolved nutrients such as nitrogen and to some extent heavy metals (Ramachandra, 2002). They are becoming extinct due to manifold reasons, including anthropogenic and natural processes. Burgeoning populations, intensified human activity, unplanned development, absence of management structures, lack of proper legislation, and lack of awareness about the vital role played by these ecosystems are the important causes that have contributed to their decline and loss. Identifying, delineating, and mapping of wetlands on a temporal scale provide an opportunity to monitor the changes, which is important for natural resource management and planning activities (Ramachandra, Kiran, & Ahalya, 2002). Temporal RS data coupled with spatial analysis helps in monitoring the status and extent of spatial features. The spectral signature associated in each pixel of the remotely sensed data is used to perform the classification and, indeed, is used as the numerical basis for categorization of various spatial features (Lillesand & Kiefer, 2002). Most of these classifications are based on certain pattern recognition techniques. Among the various frameworks in which pattern recognition has been traditionally formulated, the statistical approach has been most intensively studied and used in practice. The design of a recognition system also involves the following issues: definition of pattern classes, sensing environment, pattern representation, feature extraction and selection, cluster analysis, classifier design and learning,selection of training and test samples, and performance evaluation. Pattern recognition techniques such as neural network, decision tree, fuzzy theory, etc. have been widely used with RS data to identify the patterns in land use classes like urban, agriculture land, etc (Kwan & Cai, 1994), (Fukushima, 1998), (Gori & Scarselli, 1998) and (Lee, Liu, & Chen, 2006). One of the primary applications of pattern classification is feature extraction. Extraction of land cover features of interest from remotely sensed data leads to a number of applications for decision makers to management planners. Given an image, the classifiers can be used to categorize the image into user defined types or to identify features based on their inherent patterns. This paper focuses on identification of wetlands using unsupervised pattern classifiers.
Extracting spatial features such as wetlands (which include lakes, ponds, tanks, marshy areas, etc.) from temporal RS data helps in monitoring their status including spatial extent, physical, chemical, and ecological aspects. Traditional approaches are digitizing through visual investigation, applying a density slicing method or an edge detection method to a single band, and classification using multiple bands (Kevin & El Asmar, 1999). Among these approaches, density slicing is one of the most popular and effective methods with an appropriate threshold. However, it is often difficult to determine the threshold as in the case of Landsat Thematic Mapper (TM) band 4, if the turbidity is very high.
The utility of threshold precipitation in predicting inundation was tested with detailed field investigations to delineate the watershed and determine elevation-bounded intervals based on soil and vegetation types and densities (French, Miller, Dettling, & Carr, 2006). Landsat images in the IR bands have been used to determine the spatial change of water over a defined temporal resolution. Image subtraction followed by binary thresholding was used to extract the exact amount of change in surface water. Threshold precipitation was then compared to precipitation events that occurred within the temporal resolution of the image subtractions to determine the association between precipitation and the inundation of the lake. Near-infrared spectroscopy has been attempted by Czarnik-Matusewicz and Pilorz (2006) to monitor the properties of systems for which water is a major constituent. To distinguish flat, uniform water bodies and cloud or mountain shadows, Wilson (1997) used variance filter as a textural algorithm (TA). Output of TA was a set of rules used by a knowledge based classifier.
In this work, pattern classifiers based on unsupervised learning have been used to extract wetlands from IR Bands of temporal RS data of various spatial resolutions. The maximum likelihood (ML) estimation followed by a Bayesian classifier was employed to quantify the tradeoffs between various classification decisions using probabilistic model. This approach is useful in the smart processing and in automatic extraction of features of interest. The probabilities, mean, and covariance are estimated for the number of user specified classes and based on these a pixel is assigned to a group. The groups are then interpreted based on prior knowledge and field experience. In this method computational complexity increases with varied parametric assumptions (Duda, Hart, & Stork, 2000).
