STUDY OF EVIDENTIAL REASONING FOR IMAGE CLASSIFICATION

Abstract

Launching of a number of advanced satellite sensors and advancements in computing tools have resulted in more and more information extraction from remotely sensed data for various applications. However, information extraction from remotely sensed data involves a level of complexity. Various errors and uncertainties make the process more challenging and demanding. Development of new methods and algorithms shall always be required to improve the quality of information extraction. Over the years, a number of computationally efficient image classification algorithms have been developed to extract thematic information from the remotely sensed data. However, each has its own limitations. An effective algorithm for image classification should be able to accommodate different types of data appropriately and should be computationally efficient. In addition, it should be robust towards presence of errors in the dataset. Mathematical theory of evidential reasoning has been found to be highly effective in deriving information from input data containing errors and uncertainty. Few studies have already been reported on use of the theory for remote sensing image classification and the resulting classifier has commonly been termed as evidential reasoning classifier (ERC). The main objective of this dissertation is to_examine the efficacy of this classifier and investigate various issues in implementation of this classifier over three experimental datasets consisting of remote sensing images, NDVI layer, DEM and its derived products. For applying ERC to a dataset, the value for a parameter "bin size" needs to be provided subjectively, which is a major limitation of the classifier. In this work, we propose a new optimization algorithm based on Genetic Algorithm (GA) for bin size optimisation. The performance of this algorithm was benchmarked against sequential dependent search. The proposed GA based algorithm performed much superior to the conventional SDS algorithm, both in terms of efficiency and accuracy on all the three datasets. On an average, time taken by the proposed algorithm was only half that of the SDS algorithm. In addition, accuracy obtained for the proposed GA based algorithm was higher than the SDS algorithm. This is true for all the three datasets, indicating consistent superior performance of GA based algorithm. II The performance of GA based method and SDS algorithms was assessed when the number of input data sources in the classification process increased. It was found that as the number of bands increased the time taken by SDS algorithm increased rapidly, whereas GA demonstrated stable performance. Moreover, accuracy obtained for GA based algorithm was higher than SDS algorithm in all cases. Theoretically, it has been stated that the evidential reasoning classifier has the capability of classifying data accurately even in the presence of errors. To evaluate this postulate an experiment was framed Three types of common errors viz. missing pixels, missing lines, and random errors in remote sensing images were considered. These errors were artificially induced into the dataset and then the accuracy of classifying this dataset was found out. The results showed that ERC was much robust towards all the three types of errors investigated. For example, the overall accuracy obtained for ERC on missing pixels was much higher (61.25%) than MLC (18.82%). The results of the experiment on random errors also showed that as the percent of erroneous pixel increased, accuracy from MLC deteriorated rapidly. In contrast, ERC demonstrated a very stable performance against any increase in the percent of erroneous pixels. From these results, it may be concluded that ERC is very robust to the occurrence of any errors in the dataset and should be preferred. The theory of evidential reasoning is capable of deriving a quantified measure of uncertainty associated with the decision. This aspect was also included to develop an uncertainty map as a by-product of the image classification process. In the uncertainty map thus developed, high values were obtained for the areas in remote sensing image where clouds and shadows were present. Such uncertainty maps shall be very useful for planners and managers as these provide additional information in the form of uncertain areas in the image so that they can be more focused about the quality of remotely sensed maps. III