Research On The Theory And Method For Intelligent Extraction Of Object In Hyperspectral Imagery

Object extraction in hyperspectral remote sensing imagery is one of the key technologies inapplication of hyperspectral remote sensing, To fulfill the requirement of object extraction inhyperspectral imagery for Precision, efficiency and robustness, using the spectral and spatialinformation of hyperspectral imagery, as well as considering the characteristic of correlation inhigh dimension and non-linear separability, multi-scale extraction of object were studied in depthusing the intelligent method for machine learning and computer vision. The main works andcreations of this dissertation are listed as follows:1. A novel spectral similarity measurement which named spectral angle consine kernelmeasurement and its adaptive selection of parameter were brought forward, and applied to objectdetection and spatial neighboring clustering in hyperspectral imagery. This measurement has agood adjustability for spectral curve variation for the same materials coming from radiationintensity variations, shadow, and shading etc. The experiments were carried on with severalhyperspectral data, and the results of the experiments proved that the proposed method could notonly extends the threshold area coverage but also improves the precision of detection.2. To solve the problem of low accuracy of non spherical distribution data clustering basdon traditional clustering methods, spectral clustering was used and improved, and anunsupervised classification method for hyperspectral imagery based on spectral clustering wasgiven. The sparse affinity matrix was used by this algorithm to solve the bottleneck of RAM, andthe pretreatment of over-segmentation for hyperspectral image based on spatial neighboringclustering method was used to improve processing speed, as well as the spectral angel consinekernel measurement was used instead of the gaussian radial basis kernel function in theconstruction of affinity matrix to improve the clustering accuracy. The result of experimentsproved that, the method proposed in this paper increases the accuracy of hyperspectralunsurpervised classification.3. The positive definite kernel based on one-class support vector machine (OCSVM) wasused to object detection in hyperspectral imagery. One-class support vector machine holds theadvantages of support vector machines, and only needs the train samples of the object. Themethod brought forward in this paper selected mathematical model, designed kernel function,selected Optimal parameter, and added the theory of OCSVM into the object detection algorithmfor hyperspectral imagery. And this method improved the pricision of recognition and reducedthe demand of train samples. The experiments were carried on by several hyperspectral image,and the results of the experiments proved the validity of the method.4. The non-positive definite kernel based on one-class support vector machine object detection method was given to solve the problem of Gaussian radial basis kernel function issensitive for Euclidean distance variations of two spectral vectors, but not for the spectral curvesof a material are variety. The spectral angle cosine kernel measurement which is a non-positivekernel function was used by this method, and a proxy kernel matrix and the analytic centercutting plane method were used to computed the global optimization of non-positive one-classsupport vector machine. The analysis of experiments shows that, this method increases theaccuracy of object detection in hyperspectral imagery.5. An object extraction strategy which joint spectral and spatial information was designed.Firstly, markov random field was used to describe the spatial relations of adjacent pixels inhyperspectral imagery, and then through the MAP-MRF framework, hyperspectral imagessegmentation was transformed into a problem of minimization of energy function, finally, usingone-class support vector machine object detection result as the initial value, and graph cutmethod was used to minimize the enegy function. The contrast experiments proved that theobject extraction strategy put forward in this paper which is based on spectral and spatialinformation can efficiently improve the accuracy of object extraction in hyperspectral imagery.6. The theory of Particle Swarm Optimization (PSO) was usd to extract the endmembers inhyperspectral imagery. First, combined the local PSO with the genetic operator, a particle swarmoptimization genetic algorithm (PSOGA) for endmember selection was proposed to solve theinsufficience of traditional PSO that it could not solve the large scale optimization problems indiscrete solution space such like the problem of endmeber selection. Second, global PSO wasused to combine the two criterions for endmember extraction that are unmixing residual andvolume of simplex, and a particle swarm optimization-based endmembers estimation (PSOBEE)algorithms was brought forward. This algorithm needn’t suppose that there are pure pixels inhyperspectral images, as well as this algorithm can preserve the shape of the endmembers’spectrums. Finally, The contrast experiments proved the validity of the two types PSO-basedalgorithms for endmember extraction.