Study On Dimensionality Reduction For Hyperspectral Remote Sensing Imagery

With the rapid development of imaging spectrometers and remote sensing technology,the application of hyperspectral remote sensing images is becoming more and more widespread.However,due to its high spectral resolution and large number of bands,it brings difficulties to hyperspectral image processing,such as image classification,which produces dimensional disaster,redundant information and long running time.Therefore,the dimensionality reduction of hyperspectral image data has become one of the important contents of hyperspectral remote sensing data processing.This paper studies the technology of reducing dimension of hyperspectral images.The main works are as follows:(1)The expression and characteristics of hyperspectral remote sensing data are analysed deeply.Two methods of feature extraction and band selection for the dimension reduction of hyperspectral images are compared.The evaluation methods of dimensionality reduction are analyzed from the three aspects of information,relevance and separability.This paper mainly uses the classification accuracy evaluation method that can be considered from the perspective of separability to evaluate the effectiveness of dimensionality reduction methods.(2)The dimension reduction methods based on band selection is focused on study and the six common band selection algorithms are introduced,including information-based ABS and MI algorithm,ranking-based ID and FSD algorithm,clustering-based WaLuDi and K-Means algorithm.The common algorithm is realized by MATLAB.Through experimental comparison,it is shown that WaLuDi and K-means algorithms have the best performance,the highest classification accuracy and the weaker correlation,but the information is relatively small.The ABS algorithm has poor performance,the lowest classification accuracy and strong correlation,but the information is the largest.(3)On the basis of summing up common band selection methods,a method based on combination of clustering and the information is proposed.First,the K-means algorithm is used to divide the bands of the original data.Then,the improved ABS algorithm is used to select the representative bands as the best band combination in the clustering results.Four kinds of similarity measures are used in clustering,and four improved methods are obtained,called KABS_ED,KABS_SAM,KABS_SCM and KABS_SID,respectively.The improved algorithm is realized by MATLAB,and the improved methods of the four similarity measures are compared with experiments,and the improving method of the most suitable.the similarity measure is KABS_SCM.(4)Using experimental data verify the validity of the improved method,we obtained the following results: compared with the traditional K-means,KABS_SCM and ABS,it shows that the proposed method is better than the former algorithm;compared with KABS_SCM and other four kinds of common algorithms,it shows that the proposed algorithm has better performance than common band selection algorithms.