Robust Analysis Of Kernel Method In Information Extraction Of Remote Sensing Images

Hyperspectral remote sensing has been developed in hundreds of contiguous bands and could provide reach spectral information for accurate classification and detection. The classification algorithm and detection algorithm based on kernel method has very strong recognition ability. In recent years, it has been widely used in the field of remote sensing information extraction. But due to the high dimensionality of hyperspectral data and the large computational complexity associated to algorithm based on kernel method, limiting the application of these algorithms in the field of onboard hyperspectral data processing.Approximate computing utilizing the robust of algorithm aims at translating the precision or output quality into power saving. It can be used to reduce the power consumption of algorithms based on kernel method, so that these algorithms can be applied in the field of onboard hyperspectral data processing. In this paper, the robust of remote sensing image information extraction algorithms based on kernel method are analyzed both on data level and algorithm level. It demonstrates the feasibility of approximate computing utilizing the robust of algorithm.1.Data level robust analysis. Normalizing the original hyperspectral data and then normalized data is converted into binary system, and then the random number is injected in the least significant bits(LSBs) of the selected data. Convert the new data form binary into decimal system and evaluate the result by using the new “corrupted” data.2.Algorithm level robust analysis. Gaussian function is used to generate Gaussian noise, then the Gaussian noise is injected into the intermediate results of the algorithm, and we further evaluate the impact on the result. In this experiment, the mean value of Gaussian function is a constant zero, increasing the variance of Gaussian function to introduce more noise.In this paper, the robust of SVM algorithm and KRX algorithm are analyzed. Experimental results show that real hyperspectral image information extraction algorithms exhibit a large amount of robust. If the noise is limited to a certain range, the noise has little effect on the output of algorithm. The robust properties are the most important features when evaluating the feasibility of an approximate computation approach. It demonstrates the feasibility of approximate computing utilizing the robust of algorithm.