Parallel Construction of Global Pyramid for Large Remote Sensing Images

The pyramid model is the basis for visualization and processing of massive remote sensing data in a virtual globe system. It produces large amounts of data tiles in the process of building pyramids however, and the traditional serial CPU algorithm has difficulties satisfying requests for large images. This paper proposes a parallel global pyramid approach that exploits GPU capabilities and multi-threading to enable memory-intensive and computation-intensive tasks. A two-level decomposition strategy was developed to migrate the performance bottlenecks in data transfers among the GPU, CPU, and disk. A multi-threading strategy was used to reduce the delay between CPU and disk, and pin-memory for GPU global delay. Finally, the GPU is used to complete large-scale parallel resample computing and constructing an octree structure used to maximize reuse of data in the graphic memory. Experimental results show that the proposed method can significantly improve pyramid construction speed for large remote sensing images.