基于POS同步的PHI高光谱遥感图像几何校正

发布时间：2018-08-24 15:47

【摘要】：在机载高光谱遥感数据预处理中,高光谱图像几何校正始终是不可或缺的重要部分。依据传统算法,计算中必然性的需要巨大搜索量或者其他大量计算,要花费大量时间,而本文便是针对此因素对其算法进行优化,以达到大幅度提高数据处理速度,节约时间的目的。 Applanix公司的POS/AV位置姿态系统是用于机载传感器的POS系统,位置精度可达5cm-30cm,方向精度可以精确到二十秒至三十秒之间,从而为数据处理提供相对高精度的位置和姿态数据。本文便是充分利用高精度匹配位置姿态数据,如经纬度、海拔、侧滚角和俯仰角等参数,使用几何校正算法重采样每个像元的地面实际坐标和灰度值,从而消除遥感平台带来的几何畸变。本文首先依次介绍了高光谱遥感的特点与应用、几何畸变原因、几何校正并行的需求,其中对几何畸变影响最大的便是机载平台的不稳定,所以取得高精度的位置姿态数据尤为重要。在第二章中便详细介绍了POS系统四个组成部分的工作原理,特别阐述了GPS/IMU组合后能得到高精度位置姿态数据的缘由。同时在几何校正之前,针对高光谱图像中出现的图像行缺失、行重复误差,经Matlab仿真分析后,得出了行号值乱码规律,而且分析出此内部畸变原因除传感器误差之外,行光谱信息传输限制也是一大因素,随之提出了可行方案,并实现了图像的复原,复原后整体效果良好。然后对几何校正算法中坐标变换和重采样做了详细介绍。针对传统重采样方法,本文提出了新的重采样算法即局部搜索距离倒数法,此算法不需要大量搜索且重采样效率较高,经此算法处理后,图像校正后整体效果良好。最后针对高光谱遥感数据以波段排列的特性和单机多核的特点,提出了多线程并行方案。在坐标变换过程中,数据读写始终频繁,而数据读写是单线程进行,所以对坐标变换部分进行数据并行是不可行的,而重采样部分在计算中以波段、行数和列数为迭代次数进行计算,所以重采样适合多线程并行。在并行计算中,基于负载均衡提出了内核动态绑定模式的并行模式,然后对其重采样效率的提高做了详细的分析。最后结果表明,多线程内核绑定模式不仅大幅度提高了重采样效率,而且也大大减少了因负载均衡而带来的内存开销。随着高光谱图像分辨率的提高和多核技术的飞速发展,如何充分利用高性能计算对提高遥感图像几何校正重采样效率有着一定的研究意义。
[Abstract]:Geometric correction of hyperspectral images is always an important part in airborne hyperspectral remote sensing data preprocessing. According to the traditional algorithm, the necessity of computing needs a large amount of search or other large amount of computation, and it takes a lot of time, and this paper optimizes the algorithm to achieve a significant increase in the speed of data processing. Applanix's POS/AV position and attitude system is a POS system for airborne sensors, with a position accuracy of 5 cm to 30 cm and a directional accuracy of between 20 seconds and 30 seconds. Thus, the position and attitude data are relatively high precision for data processing. In this paper, high precision matching position and attitude data, such as longitude and latitude, altitude, roll angle and pitch angle, are fully used to resample the actual ground coordinates and gray values of each pixel by using geometric correction algorithm. Thus the geometric distortion brought by remote sensing platform is eliminated. In this paper, the characteristics and applications of hyperspectral remote sensing, the causes of geometric distortion and the requirement of parallel geometric correction are introduced in turn. The instability of airborne platform is the most important influence on geometric distortion. So it is very important to obtain the position and attitude data with high accuracy. In the second chapter, the working principle of the four components of POS system is introduced in detail, especially the reason why the high precision position and attitude data can be obtained by GPS/IMU combination. At the same time, before geometric correction, aiming at the line missing and line repeating error in hyperspectral image, after Matlab simulation analysis, the law of line number disorder code is obtained, and the reason of this internal distortion is analyzed besides sensor error. The limitation of line spectral information transmission is also a major factor. A feasible scheme is proposed and the image restoration is realized. The overall effect of the restoration is good. Then, coordinate transformation and resampling in geometric correction algorithm are introduced in detail. For the traditional resampling method, a new resampling algorithm, the local search distance reciprocal method, is proposed in this paper. The algorithm does not require a lot of search and the resampling efficiency is high. After the algorithm is processed, the overall effect of image correction is good. Finally, according to the characteristics of hyperspectral remote sensing data arranged in band and single machine multi-core, a multi-thread parallel scheme is proposed. In the process of coordinate transformation, data reading and writing is always frequent, and data reading and writing is carried out by a single thread, so it is not feasible to parallelize the coordinate transformation part, while the resampling part takes the band in the calculation. The number of rows and columns is calculated by iterations, so resampling is suitable for multithread parallelism. In parallel computing, the parallel mode of kernel dynamic binding mode is proposed based on load balancing, and the improvement of resampling efficiency is analyzed in detail. The results show that the multi-thread kernel binding mode not only improves the efficiency of resampling greatly, but also reduces the memory overhead caused by load balancing. With the improvement of the resolution of hyperspectral images and the rapid development of multi-core technology, how to make full use of high performance computing has a certain significance in improving the efficiency of geometric correction resampling of remote sensing images.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】