A Hybrid Process/Thread Parallel Algorithm for Generating DEM from LiDAR Points

Abstract

Airborne Light Detection and Ranging (LiDAR) is widely used in digital elevation model (DEM) generation. However, the very large volume of LiDAR datasets brings a great challenge for the traditional serial algorithm. Using parallel computing to accelerate the efficiency of DEM generation from LiDAR points has been a hot topic in parallel geo-computing. Generally, most of the existing parallel algorithms running on high-performance clusters (HPC) were in process-paralleling mode, with a static scheduling strategy. The static strategy would not respond dynamically according to the computation progress, leading to load unbalancing. Additionally, because each process has independent memory space, the cost of dealing with boundary problems increases obviously with the increase in the number of processes. Actually, these two problems can have a significant influence on the efficiency of DEM generation for larger datasets, especially for those of irregular shapes. Thus, to solve these problems, we combined the advantages of process-paralleling with the advantages of thread-paralleling, forming a new idea: using process-paralleling to achieve a flexible schedule and scalable computation, using thread-paralleling inside the process to reduce boundary problems. Therefore, we proposed a hybrid process/thread parallel algorithm for generating DEM from LiDAR points. Firstly, at the process level, we designed a parallel method (PPDB) to accelerate the partitioning of LiDAR points. We also proposed a new dynamic scheduling strategy to achieve better load balancing. Secondly, at the thread level, we designed an asynchronous parallel strategy to hide the cost of LiDAR points’ reading. Lastly, we tested our algorithm with three LiDAR datasets. Experiments showed that our parallel algorithm had no influence on the accuracy of the resultant DEM. At the same time, our algorithm reduced the conversion time from 112,486 s to 2342 s when we used the largest dataset (150 GB). The PPDB was parallelizable and the new dynamic scheduling strategy achieved a better load balancing. Furthermore, the asynchronous parallel strategy reduced the impact of LiDAR points reading. When compared with the traditional process-paralleling algorithm, the hybrid process/thread parallel algorithm improved the conversion efficiency by 30%.