Spatial Statistics Based Hyperspectral Band Selection After Dimensionality Reduction

Hyperspectral sensors obtain detailed spectral characteristics of the target on the ground, greatly improving the ability to identify the features.It has been widely used in geology, environment and disaster, fine agriculture, geomaticsand archaeological, etc. It is difficult to process and imagery interpretbecause of the numerousbands, redundance and large amount of hyperspectral remote sensing image. The classification accuracy of the hyperspectral remote sensing image will be improvement as the number of the bands increase, and then decrease, we call it“Hughes”phenomenon. There are up to 94% of spectral band are not necessary based on the research while the classification accuracyhold the line. Therefore, we need to reduce the dimensions of the hyperspectral images firstly in the application. There are many researches about hyperspectral image dimensionality reduction algorithm so far. The linear dimensionality reductionalgorithms are mature,such as principal component analysis, the maximum noise fraction transform and independent component analysis, has been used as a module of the software ENVI, etc.Nevertheless, the researchabout the band selection afterdimensionality reduction are less at present. Traditional wavelength selection methods after dimension reduction, for example, cumulative characteristic values(or variance), signal-to-noise ratio and negative entropy, are unreasonable to determine the number of wavelengths, not applied to all of the dimension dimensionality reduction, and not taken into account the relationship between image features and spatial position, not a very good to explain the threshold.Spatial statistics iscommonly used to describe the spatiality and randomness of the regionalized variable. Therefore, this study explore the reasonable method of band selection after dimension reductionbased on spatial statistics, which is used to different dimensionality reductionalgorithm and determined the experience threshold.Fitting the theory semi variation function automatically by increasing the experiment semi variation function points gradually with the cross validation method.The calculation of semi variation function and fitting is the fundamental problem of spatial statistics. The semi variation function can be achieved conveniently by ArcGIS, GS+, etc. We need to calculate and fitsemi variation functionautomatic because it will be very inconvenient as the number increasing of the hyperspectral remote sensing image and the Artificial solving efficiency is very low.Two methods based on theory semi variation functionparametersand fractal dimension are proposed to determine the bands after dimensionality reduction. We select the bands after dimensionality reduction based on the range with arch rise/sill from theory semi variation function and fractal dimension from the experiment semi variation function, which is be different from the traditional methods. The bands are filtered by the classification accuracy of the mahalanobis distance classification algorithm, and finally we determined the band selection scheme and the threshold after dimensionality reduction.Finally, the range is 4.5, arch rise/sill is 0.2 and absolute value of the fractal dimension minus two is 0.02 by the experiment and.