Research On Texture Classification In Remote Sensing Image With Wavelet-based Hidden Markov Tree Models

Remote sensing image has rich texture information, and exact texture feature extraction is a key factor to both image segmentation and classification. Model-based method is a basic method to the texture feature extraction. It is also a method appropriate to remote sensing image texture analysis.Considering the multi-scale characteristic and the random feature of the remote sensing image, the wavelet-based HMT model is adapted to analyze the texture characteristic. Further research is carried out according to the requirement and characteristic of remote sensing image. Because the wavelet coefficients don't accord with the Gaussian distributions and there are underlying relationships among these wavelet coefficients on both the same scale and the inter-scales, the wavelet-domain HMT model reveals exactly the dependencies among these coefficients. It estimates the model parameters using the expectation maximization (EM) algorithm and processes image classification using Marxism Likelihood (ML) method.Concerning with the characteristics of remote sensing image, based on the description of the basic model establishment and segmentation, further study is conducted in this thesis on the wavelet-domain HMT model itself and its application in remote sensing images. It studies and resolves the problem of the non- consistent size of the training sample with the original image. A method is presented to segment image with once model establishing, multi-layer segmentation results getting at the same time. Aiming at the multi-spectral remote sensing image, the spectral information is fused through HIS transform or PCA (Principle Component Analysis) ,from which we can get the description of different aspects in image. The HMT is then applied to the fused image to get the new segmentation result. The experimental results show that, it helps improving the segmentation result when appropriately using the multi-spectral information. Besides, this thesis deeply studies the multi-scale effect of HMT-based texture description and some related problems. The results may be of use to the future research work.