POPS: An Efficient Framework for GPU-based Feature Extraction of Massive Gridded Planetary LiDAR Data [Scalable Data Science]
AbstractOne of the long-standing question in geoscience is whether there is a topographical signature of life. Recent development of space-borne LiDAR has led to massive data depicting planetary topographies, opening up unprecedented opportunities to make progress in answering this question. This is what we set out to do in an ongoing project named í µí±í µí°´í µí±í µí°¾í µí°¸í µí±. A key step of PARKER is to find topograph-ical features that are potentially relevant to intrinsic differences between Earth and alien worlds. Due to the huge data volume, sequential feature extraction cannot meet our needs in PARKER. Hence, in this work we propose a GPU-accelerated framework for fast feature extraction of planetary LiDAR, which as far as we know is the first GPU-based solution for this task. Faced with multi-scale features and limited GPU memory, we present a novel pseudo-one-pass sweep (POPS) approach, leveraging memory-aware data grouping and incremental data transfer to address these challenges. We also develop a GPU-based solution to aggregate features extracted by POPS. Experiments on real and simulated data show that our algorithms are 2-3 orders of magnitude faster than their sequential counterparts and 1-2 orders of magnitude faster than MPI-based multi-core parallelism, enabling near real-time analytics of datasets with almost a billion points. Based on POPS, we have been able to efficiently evaluate the relevance of topographical features to intrinsic inter-planetary differences. So far, we have assessed the abilities of two feature extractions methods, PCA and STAT, to capture differences between Earth and Mars. Results show that PCA features on scales of 300-500m can best capture such differences. Thanks to the generic nature of POPS, we will be able to expand our studies to new feature extraction methods and other alien worlds than Mars in the next phase of PARKER.