The Application Of High Performance Computing In Hyperspectral Remote Sensing Data

Hyper-spectral remote sensing technology is the forefront of remote sensingtechnology; it contains a wealth of spectral data which makes Feature quantitativedetection possible. At the302th Xiangshan Science Conference which main issue isthe new challenge faced by remote sensing prospecting. They agreed thathyper-spectral remote sensing technology enhance the capacity of remote sensing inearth observation and the ability to identify, it greatly improves the capacity ofquantitative processing in remote sensing technology. Hyper-spectral remote sensingtechnology enables the development of remote sensing from the identification ofsurface features to the feature identification phase, using hyper-spectral remotesensing data for mineral exploration is one of the main directions of the application.The remote sensing technology has been widely used in resource exploration, disastermonitoring and environmental monitoring. With the implementation of China’s "highscore" and success of launch and in orbit of Resource One02C satellite.Hyper-spectral data source for the bottleneck problem will be solved step by step.Hyper-spectral remote sensing data includes information on space radiation andspectral triple, and the amount of data in industrial applications has reached to TBlevel, mass characteristics of the data has seriously hampered the efficiency ofapplication development and practical application. Such as size30000*30000satellite remote sensing image, if use the traditional serial pounds correction, theoperation will reach to tens of billions of floating-point multiply-add operation; alarge number of data manipulation and processing complexity decide the image dataprocessing of the hyper-spectral remote sensing has a strong computing, ordinarycomputers and expensive dedicated hardware system far unable to meet the growing demand of remote sensing data processing. In addition, the application such as imagepreprocessing, correction, classification, mapping algorithm of hyper-spectral remotesensing is very complex, these factors have seriously hampered the efficiency ofmineral exploration work in the actual industrial applications. High-performancecomputing is the development of computer technology, but also the low effective wayto solve the efficiency of mass data processing, it has been used in many computingarea which has a large amount of data, but for the mass of high-performancecomputing of hyper-spectral remote sensing data.Because of hyper-spectral remotesensing data corresponding to hundreds or even thousands of bands ranging from data,data has a correlation between band spectrum-dimensional data and spatial dimensiondata integration, parallel processing, the simple pursuit of calculation speed in theapplication side but will result in constraints. Besides this, due to the limitations ofcomputer memory, read huge amounts of data sub-block operation must take intoaccount the uniform between spectral dimension and space dimension. This paperresearched the parallel environment hyper-spectral remote sensing high-performancecomputing optimization strategy for the mass hyper-spectral remote sensing dataprocessing and parallel environment suitable for CUDA (Compute Unified DeviceArchitecture) and MPI (Message Passing Interface), whose parallel processing modelis used for hyper-spectral remote sensing image characteristics. Hyper-spectral remotesensing decomposition of mixed pixels and dimensionality reduction processingalgorithms, for example, corresponding decomposition treatment of high-performanceparallel algorithm rewritten, were constructed based on CUDA and MPI paralleldevelopment of architecture, design and development of high-performancehyper-spectral remote sensing mixed algorithms and procedures, integrated computersoftware and processing of hyper-spectral remote sensing technology, hyper-spectralremote sensing approach is to further develop high-performance computing researchfor the sustainable socio-economic development and mineral resources explorationwork to provide technical support. Finally, based on the theoretical studies the dataprocessing of the CPU/GPU hybrid parallel mode, establish a data calculating modelunder the hybrid parallel environment.1) Massive hyperspectral remote sensing data model for two parallelenvironments.2) Study the huge amount of read-write algorithm data, breaking the protectionof commercial software on the the mass data source extraction method source code. 3) Research and design the paraller mode of hyperspectral remote sensing-basedGPU and mixed pixel decomposition the approach parallel algorithm, for exampleprogram implementation and testing, including: unconstrained least squares, andconstrained least squares method1, endmember projection vector method.4) study and design a parallel algorithm based on MPI parallel modehyperspectral remote sensing dynasties dimension, for example the verificationalgorithms, including principal component analysis (PCA) and minimum noisefraction (MNF).5) For CUDA+MPI mixed parallel environment under algorithm design anddata model theoretical research, laid a solid foundation for the depth ofhigh-performance computing applications in hyperspectral remote sensing.