| Classification and pattern recognition of high dimensional remote sensing data is distinctly different from traditional multi-channel remote sensing classification techniques. Regarding to the abrupt increase of spectral dimensionality, far more training samples than traditional multi-spectral classification are needed to accurately estimate those statistical variables which are used to describe pattern properties existed in feature space. Hughes phenomena shows that feature extraction and preprocessing before classification of high dimensional remote sensing data are theoretically necessary and effective in practice. Thus, the exploitation of reliable and efficient feature extraction techniques for high dimensional data has been one of the most desirable research topics in remote sensing pattern classification and recognition fields.This dissertation investigates issues on methodology and applications of feature extraction and image classification for high dimensional remote sensing data. We analyzes and discusses the main accuracy influences existed in image feature representation, data preprocessing, feature extraction, spatial feature representation and processing, image classification algorithm in a scenario of land cover classification and recognition based on high dimensional remote sensing data. We also apply feature extraction techniques on practical issues, in which we discusses the application of NOAA-AVHRR 1km time series data for land cover classification in China.Main research topics and initiatives in this thesis include:1. An improved spatial domain image deblurring algorithm is proposed based on wiener filter and deconvolution.Through analysis the concept of point spread function (PSF) for high resolution remote sensing imagery, we discuss estimation of the synthesized point spread function and image restoration algorithm. This PSF estimation algorithm could be applied on CBERS-1 satellite visible and infra-red images in different atmospheric conditions without suffering restoration performance. We use wiener filter to compute the deconvolution operator. The quality of the imagery restored from the deconvolution operator is evidentially improved. Moreover, since the restoration procedure is performed in spatial domain, it is relatively simple and low computational complexity.2. A non-linear adaptive filter for impulse noises removal been proposed in this thesis.The key point of the algorithm is to sort the pixel values, to compute the sequences of standard deviations and means, and then to take the normalized differences between two successive standard deviation/mean ratios. Noise detection is achieved by thresholding these differences. Noise suppression is achieved by replacing the pixel value with the rank-ordered mean. The algorithm has been tested on simulated data and real SeaWiFS image so that its superiority is established. By considering the generalized impulse noise model discussed in the paper, we analyzed the quantitative relationship between the noise removal performance and the threshold parameters. Finally, the strategy for retrieval of the optimized thresholds has also been presented for non-recursive implementation.3. A new feature selection technique based on genetic algorithm is proposed for selection of optimal high dimensional subset data using classification training samples.The main popular feature subset selection technologies are reviewed in this thesis. We have proposed a new feature selection technique based on genetic algorithm (GA). This algorithm could heuristically search the optimal feature subset from different classification band combinations. Preliminary maximum likelihood classification results show that the proposed genetic algorithm is more probably to find the best feature subset.4. A new feature extraction algorithm based on GA and wavelet transform is proposed for high dimensional data reduction and classification accuracy improvement.The main popular feature extraction technologies are reviewed in this thesis.
|
PDF Full Text Request