Research On Techniches For Mixel Classification Of Multispectral Imagery

Several key problems in mixel classification of multispectral imagery are studied in this dissertation, which include determination of the number of endmembers, extraction of initial endmembers, speeding extraction of imagery endmembers, mixel classification method with high precision, etc. The main and original works are as follows:1. In the dissertation, we first analyze the background and current research of the two key problems of mixel classification, which are endmember extraction and abundance estimation, and point out the main problems to be studied. Then, vital theory and methods are put forward, which are the base of the study of this dissertation.2. Theory and methods of endmember extraction and abundance estimation are studied. Then we analyze the physical mechanism of linear and nonlinear mixed spectrum, and the applicability of linear and nonlinier classification models, and accuracy assessment principles of classified results. The relationship between the simplex which is the data cloud of imagery in its feature space and endmembers is sdudied. Then, we compare quantificationally the performance of four kinds of Endmember Extraction Algorithms on precision, efficiency, anti-noise, sensitivity to outlier in a great deal of experiments, in which what factors can deteriorate the precision of classification is also discussed. At last, several important conclusions are achieved.3. The automatic methods of determinating the number of endmembers are sdudied. Then, a new method based on endmember independency is put forward. Experiment results show that not only the number of endmembers can be determinated by our method correctly, but also the efficiency of the algorithm is higher.4. The influence of initial endmembers and the strategy of searching endmember on the stability of endmember extraction algorithm is discussed. And the approaches of settling these problems are analyzed. Then, we put bring forward a new strategy of searching endmember, which can reduce the algorithm complexity, and improve the precision of endmembers to be extracted.5. The character of simplex boundary is studied. And the problem of extracting simplex boundary is settled by replacing simplex boundary with 2-D scatter boundary. The relationship between simplex boundary and endmembers is sdudied, and then a new endmember extraction algorithm (EEA) with higher efficiency, which is referred to as Simplex Boundary Algorithm (SBA), is achieved. Experiment results show that the anticipant endmembers can be extracted by SBA, and the endmember search scope can be shrinked greatly, thus the rate of endmember extraction can be increased acutely.6. The utilization of binary decision tree in classification of remote sensing imagery is discussed. A new mixel classification algorithm based on binary decision tree (BDT) is put forward by introducing BDT into mixel classification, and establishing reasonable decision judgement principal. Studies show that our algorithm can reduce the influence of endmember relativity on classified results, thus improve the precision of classified results.