Research On Remote Sensing Image Normalized Difference Vegetation Index Based On GPU

The data resolution of remote sensing image is higher and higher when the technology of getting the data is becoming mature and the way also is diverse. So, the amount of image data grows exponentially and the size of the single image is doubled and redoubled. The cell processor and memory bandwidth which are used to process remote sensing image have a higher request. Eventually, the load and running time increases sharply when computer processes the remote sensing image. The Systems of remote sensing image processing often need to have the ability to deal with mass data on time and in real time. The method of remote sensing NDVI based CPU serial architecture already can't meet the demand. So far, the methods mainly focused on the use of multi-core CPU to parallel process or part of the program achieve accelerated processing via the GPU,but the whole of calculation process don't have strict task partitioning and also don't design reasonable plan about threads allocation when people want to extract the NDVI information from remote sensing image. So, it is particularly important to research a parallel algorithm that can improve the execution efficiency of the NDVI. In order to solve the above problems, a method of remote sensing NDVI based on GPU parallel architecture is put forward in this paper. On the one hand, the author designs the parallel algorithm of the process in detail and divides the computing tasks in a reasonable manner which will do parallelization for some works according to the characteristics of the GPU. On the other hand, the author also works out the optimal thread mapping model and researches the specific technology which is used to optimize the GPU program. Specifically, the main work and contributions are the following:1. This article simply introduces the history and development trend of the GPU and discusses CUDA technology foundation.2. Through the analysis of the NDVI research content, the author uses the ERDAS IMAGINE to achieve the band separation for multispectral remote sensing image and obtains the experiment data, such as the infrared spectrum and near infrared band. Image read and display based on the Open CV. After image data into memory, will be stored in the Mat format. In this paper, a detailed discussion of the remote sensing image processing operations mainly include histogram calculating, histogram extension, histogram equalization and image binaryzation so that determine whether the algorithm can be parallelized.3. Firstly, it is necessary that research parallelization feasibility for this algorithm. Secondly, the author designs the flow of the parallel algorithm and researches the principle of task partitioning about CPU and GPU. The kernel function's threading model is organized in the form of two-dimensional mapping. What is more, GPU memory takes advantages of shared memory, which is faster to read and write compared with the global memory. In order to avoid causing bank conflict from shared memory, we solved the problem by thread.Dim.x+1 in line. Lastly, the parameters for execution of the kernel function are optimized by hardware environment, namely the number of block and thread so that make the GPU performance to the maximum.4. The algorithm is proved right by a series of tests. We can draw a conclusion that the method of remote sensing NDVI based GPU can obtain a great speedup on the application.