Application Of Non-negative Matrix Factorization Metho In Remote Sensing Image Recognition

Remote sensing information has been one of the important parameters to our understanding of nature and human ecology system of the spatial distribution. With the continuous development of remote sensing technology, and the unceasing enhancement remote sensing data of radiation resolution, our ability to extract from remote sensing image data of quantitative biophysical parameters, the digital image processing and land use (land cover information), is also greatly enhanced.The traditional remote sensing image recognition methods, such as nonparametric methods in identifying and calculation, do not require any empirical model on the intervention. But the nonparametric method has the main problem: even if the user master some of the characteristics of data, or have some prior knowledge of the observed object, they still cannot intervene in the process, consequently they can not obtain the desired results and the processing efficiency is also not high. The another drawback of parametric processing is that the computation usually involves SVD (singular value decomposition) and a large amount of data, that makes the process running slowly.The NMF (nonnegative matrix decomposition) method can overcome the above disadvantages to some extent: it provides a new idea of matrix decomposition and has clear physical meaning due to non-negative constraints. NMF saves a lot of storage space by avoiding the SVD and the eigen-computations. Through nonnegative matrix decomposition, NMF generates two factor matrices, the basic matrix and the code matrix, reducing the dimensions of the large data matrix, and as a result makes data storage capacity greatly compressed. Thus NMF can speed up data processing, and can reflect the internal structure and characteristics of the given dataset.In this thesis, we use both the of nonnegative matrix theory and Matlab programming to resolve or improve the remote sensing image processing results in existing pattern recognition, step by step. We first investigate the characteristics of the given nonnegative data matrix itself, and then analyze the composite structure of the two factors produced by NMF by adding the restriction of the sparseness, and then consider the relationship between the factors matrix and the original matrix, thereby mining the connection between the code matrix and the original image to improve existing NMF algorithm.