| With the increase of multi-band and multi-resolution remote sensing data, more effective algorithms are needed to improve the classification precision of the new remotely sensing data. Traditional supervised classification methods require the assumption that data distribution obeys Gauss normal distribution, however, actual remote sensing data distributions do not always obey Gauss normal distribution, which is one of the reasons of low precise classification by using traditional methods. Therefore, the perspective of the dissertation is to select some nonlinear method and to achieve combined results. The basic elements used in combined methods are Neural Network Classification, Wavelet Transformation, Fuzzy Mathematics, and Rough Sets. The algorithms are realized by developing new software.In the first chapter, developments of remote sensing image classification neural network theory and present research results are introduced, then the necessity of the combination of neural network classification method and other non-linear mathematic methods are analyzed. In the second chapter the structure, learning rule, and main types of neural network are introduced. In the third chapter, firstly, BP neural network algorithm is introduced, Secondly, base on the BP neural network algorithm principal, the classification of the training data collected in weather stations locally, and performed in AVHRR thermal data the test is successfully in this way the station data can be expansion to larger scale. In the fourth Chapter, Rough Sets theory is used in image data processing. In particular, the allowance rough sets theory developed from 1990's is introduced. The allowance rough sets can be used as preprocessing procedures before performing BP neural network algorithm, in this way the problems of BP neural network convergence is well proved. In the fifth chapter, at first the Wavelet Transformation theory is introduced and used in fusion multi band satellite data. And then combined with the Self-organizing Feature Map Neural Network, in the procedure the vector quantification learning algorithms are presented. In the sixthchapter, mainly dealing with clustering method after analysis are compared of the K-mean algorithm and Fuzzy C-mean algorithm in clustering. A new improved fuzzy c-mean algorithm is developed and results are proved using M ellipse measurement. Finally some discussion are made in the clustering centers can be used as the centers needed by RBF neural network for next step research. In the seventh chapter, conclusions and discussions are made.Some algorithm innovations of the papers are listed as following,1. In this paper, the rough sets theory and its application are analyzed and tested in processing remotely sensed data.2. In this dissertation particularly the allowance rough sets is used as preprocessing methods for BP neural network algorithm, in this way some noise data are removed and the problems of BP neural network convergence is well proved.3. The theory of Self-organizing Feature Map Neural Network Algorithm (SOFM) is introduced and performed by using combination of the Wavelet algorithm. During the procedure the SOFM maps multi-dimension spatial data into low-dimension space without changing their inner topological features. Moreover, the visible effect of classification can be easily judged in the topological map in low-dimension space.Some merits are presented as following,4. Normally, the same spatial data source is used as train data and to use different data source as training data have some advantages in scale expending. The BP neural network used in this aspect is tested and introduced in the dissertation.5. A Wavelet Transformation Algorithm developed in data fusion. The spatial resolution of ETM+ multi bands are dramatically proved through local high frequency substitution of Pan data.6. The Mahalanobis, the ellipse spheroid cluster algorithm, measurement is used in fuzzy c-mean clustering method, which improves the clustering accuracy comparing w...
|
PDF Full Text Request