Abstract
This thesis is concerned with the problem of mixed pixel classification in Remote sensing applications and attempts to find accurate and robust solutions to this problem. The application we are interested in, is to monitor burned forest regions for a few years after the fire in order to identify the type of vegetation present in these areas and consequently assess the danger of desertification. The areas of interest are semi-arid where the vegetation tends to vary at smaller scales than the area covered by a single Landsat TM pixel, thus mixed pixels are quite common. In this thesis we considered whole sets of mixed pixels. First, an overview of the methods currently used to solve the mixed pixel classification problem is presented, focused on the linear mixing model which is adopted in this thesis. Then a method that incorporates higher order moments of the distributions of the pure and the mixed classes is proposed. This method is shown to augment the number of equations used for the classification and theoretically it allows the specification of more cover classes than there are bands available, without compromising the accuracy of the results. The problem of deterioration of the classification performance, due to inaccuracies in calculation of the statistics when outliers are present, is also examined. The use of the Hough Transform is proposed for the linear unmixing in order to provide robust estimates even in cases where outliers are present. The Hough transform method though, is an exhaustive method and therefore has higher computational complexity. Furthermore, its performance, in the absence of outliers, is not as good as the solution obtained by the Least Squares Error method. Hence, the Randomized Hough Transform is proposed in order to improve the computational speed and maintain the same level of performance, while the Hypothesis Testing Hough Transform is proposed to improve the accuracy of the classification results. All the methods proposed in this thesis have been compared with the Least Squares Error method using simulated and real Landsat TM image data, in order to illustrate the validity and usefulness of the proposed algorithms.