Improvement and Estimation of Classification Accuracy for Remotely Sensed Images

The aim of the study in this thesis is to improve the reliability of the land use inventory, by focusing on improving and estimating the classification accuracy of remotely sensed images.;Firstly, a solution to unmix mixed pixels for hyper-spectral images is proposed. One of our contributions is the proof for AMEE assumption and the proposed Improved AMEE to enhance the performance of endmember extraction. Pre judgment before applying the Least Squares is carried out to calculate the abundances. To further improve the performance of pixel unmixing, Reliability-based Sub-pixel Mapping is proposed to locate each endmembers in a mixed pixel.;Secondly, pixel unmixing for multi-spectral remotely sensed image based on single band is also addressed. Mountain Clustering is applied to extract endmembers. The Grey Correlation method is used to generate the abundance of each endmember. The Improved Cellular Automata is proposed for sub-pixel mapping. Finally, Multiband Synthesis is carried out to integrate the results from all the bands.;Thirdly, an Eigen values-based Multiple Classifier System is proposed to improve the accuracy of land use classification. A posterior probabilities matrix of each component classifier is obtained and the eigen-values are calculated based on the above matrix. These eigen-values are used to weight the classifiers. With the proper correspondence between eigen-values and classifiers, the proposed method has proved to be effective for image classification land use inventory.;Fourthly, a sampling strategy based on error-distribution is proposed to assess the accuracy of the image classification results. The error of image classification is assumed to follow the normal distribution and can be filtered by different Gaussian filters with multi-scale. Therefore, error surface is obtained by subtracted error-free data from the classification results. The extreme points on the error surface are considered as the representative points of the error so that the least sample size is determined.;In sum, this research contributes to the quality of the land use inventory as follows: (a) improve image classification accuracy by solving the mixed pixel problem and provide a multiple classifier system; and (b) improve the reliability of the accuracy assessment result by a novel spatial sample strategy.