Parallel Computing and Performance Optimisation in Remotely Sensed Image Processing

Abstract

Remote sensing is the acquisition of physical response from an object without touch or contact, often collected by remote sensors mounted on satellites or aircraft. Such datasets have been used for many decades in a wide range of applications. With advances in sensor technology, earth imaging is now possible at an unprecedented level of detail, and the amount of data acquired by imaging sensors has been growing rapidly in recent years. Many procedures for processing remotely sensed information are characterised by massively parallel data processing, intensive computation and complex processing algorithms. These characteristics make real-time processing of large datasets very crucial. Graphical Processing Unit (GPU) is a typical parallel computing and multicore architecture, designed to perform computations on large amounts of independent data. In recent years, GPU has become dominant for high performance computing and its massive computational capability is well suited to analyse large-scale remotely sensed information. The thesis studies the computing architecture and memory hierarchy of GPUs accelerators, and discusses the main performance issues in scientific and parallel computing. Parallel computing frameworks, parallelisation strategies, performance optimisation and acceleration algorithms for remotely sensed image processing are provided. Effective and efficient solutions are provided to dynamic programming based NP-hard optimisation problems and pixel-classification based hyperspectral imaging unmixing procedures. Verification methods are desgined to evaluate performance in terms of accuracy and speedup. The proposed methods assume spatial smoothness in remotely sensed images, which distinguishes them from artificial arts or natural photos. The benchmark tests show good performance accelerations compared with both sequential and parallel implementations in the literature on NVIDIA GPU accelerators. An analysis of performance profile results is presented and can be referred as a guide to similar parallelisation strategies for other computing platforms.