The Research On Adaptability Of Multi-platform Feature Extraction For Hyperspectral Image

As the hyperspectral image begins to present high time resolution, high spatial resolution, and high spectral resolution, the data volume increases massively, it makes the hyperspectral processing image more complex, more time-consuming and requires more for hardware platform to process it. That means the real time or near real-time processing of hyperspectral image becomes a big problem. So far, numerous scholars have put forward some methods to process hyperspectral image from a vertical platform, but almost all of which research from a single platform and lack resource integration, analysis, and comparison of its multiple languages and platforms. Thus, according to the characteristics of hyperspectral image from horizontal and vertical aspects, combining with its processing language and platform environment, this study is aimed at exploring the language and platform which is authentically adaptable to hyperspectral image processing, from the aspects of time consumption, speed ratio, energy consumption, complexity, as well as the results of image processing. From vertical and horizontal aspects, the purpose of this study is to find a new way of thinking to explore the adaptability of hyperspectral image processing platform. This thesis mainly focuses on the following:First of all, the programming model, memory model and processing mechanism of CUDA are analyzed, and the program performance optimization problem is discussed based on CUDA in this thesis. Next, it explores the deficiencies from the typical hyperspectral image feature extraction algorithm. And then, based on the two algorithms of parallel minimum noise fraction and the principal component analysis under GPU parallel environments of CUDA architecture, it optimizes and improves the data communication, data partition and memory access to speed up the processing of hyperspectral image feature extraction. Through the simulation experiment, the highest speed-up ratio of the two algorithms can reach 122 times after optimization. It optimizes the processing time and improves the speed-up ratio. Finally, in view of various languages and platforms to process the hyperspectral image at present, it puts forward a multi-platform mechanism(ENVI, Matlab, series and parallel environments) to explore the simulation experiment on hyperspectral image feature extraction. The simulation results are compared from the vertical aspects and the comprehensive advantages and disadvantages of various processing platforms are evaluated, in order to provide a new idea for the adaptability of hyperspectral image processing platform. It lays a foundation for the follow-up research work, such as hyperspectral image classification, target detection, decomposition of mixed pixels and so on. This also makes it possible to process hyperspectral images quickly and efficiently.