CN-122023132-A - Geographic space grid multi-resolution streaming construction method based on CPU-GPU cooperation

CN122023132ACN 122023132 ACN122023132 ACN 122023132ACN-122023132-A

Abstract

The invention discloses a geospatial grid multi-resolution streaming construction method based on CPU-GPU cooperation, which comprises the steps of preprocessing multisource geospatial data, estimating dynamic analysis data space distribution density based on kernel density by a CPU, executing self-adaptive space blocking to balance calculation load, transmitting blocking data to a GPU through an asynchronous pipeline, constructing a highest resolution basic grid by the GPU in parallel, then continuously recursively executing voxel aggregation and downsampling in the GPU, combining edge preservation constraint, a statistical moment conservation mechanism and the like to generate a complete multi-resolution grid pyramid, and finally publishing results in real time through a streaming interface. According to the invention, through the CPU-GPU deep cooperation, self-adaptive blocking and streaming processing architecture, the memory peak value and the processing delay are obviously reduced, and the ultra-large scale geographic space data meshing efficiency and quality are improved.

Inventors

ZHANG YUTIAN
LIU YUANYUAN
LIU YANG
ZHANG QIANG
YIN HUI
LI DONG
ZHU LI
QIN ANNI

Assignees

中国电子科技集团公司第十五研究所

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (10)

1. The geospatial grid multi-resolution streaming construction method based on CPU-GPU cooperation is characterized by comprising the following steps: s1, data preprocessing, namely acquiring multisource geospatial data and preprocessing the data; S2, CPU self-adaptive space blocking, namely, at a CPU end, analyzing the space distribution density of the preprocessed data by adopting a kernel density estimation algorithm, dynamically determining the blocking size according to the space distribution density, and executing self-adaptive space blocking to generate a plurality of blocking data; S3, GPU parallel gridding processing, namely transmitting each piece of block data and metadata thereof to a GPU video memory through a pipeline parallel strategy, and executing gridding calculation kernels on the GPU in parallel to calculate and obtain the original data layer LOD 0 grid data with the highest resolution; S4, continuously generating multiple resolutions of the GPU, namely acquiring LOD 1 -LOD n grid data of a downsampling layer in a recursive aggregation mode based on LOD 0 grid data in the GPU; S5, stream output and release, namely, the LOD 0 grid data and the LOD 1 -LOD n grid data are assembled into a multi-resolution grid pyramid and released in real time through a stream output interface.
2. The method of claim 1, wherein the multi-source geospatial data in S1 comprises laser point cloud data, digital elevation model data, multi-spectral remote sensing images, and the data preprocessing comprises coordinate system unification, data denoising filtering, and data format standardization.
3. The method of claim 1, wherein in S2, n data points p= { P 1 , p 2, ..., p n }, where the i-th data point P i = (x i , y i , z i ), in a given three-dimensional space, the probability density function of the kernel density estimation algorithm is: Wherein, the Is indicated at the point The probability density value at which the probability density value is calculated, Three-dimensional coordinates that are function arguments; k (·) is a kernel function, and an isotropic Gaussian kernel function is adopted; h is a bandwidth parameter, and the improved Silverman rule is adopted for self-adaptive determination: Wherein the method comprises the steps of As dimension, σ is the mean of the standard deviation of each coordinate for the data point.
4. A method according to claim 3, wherein in S2, when determining the block size: the mapping relation between the block size and the local density is determined by a nonlinear function, and the formula is as follows: Wherein S represents the block size, d represents the local density, 、 To preset the minimum and maximum block sizes, 、 Normalized density values for minimum and maximum; parameters for controlling the variation of the block size along with the density; parameters for controlling the curvature of the function; according to the block size function, adopting an octree data structure to realize space recursion division on the whole; Based on spatial adjacency, density similarity and load balancing, adjacent blocks are combined by adopting a condensation hierarchical clustering algorithm.
5. The method of claim 1, wherein in S3, a multi-CUDA stream parallel transmission mode is adopted, and a double buffering mechanism and a page lock memory are combined to optimize a data transmission bandwidth.
6. The method of claim 1, wherein in S4, a hierarchical parallel downsampling strategy is adopted during recursive aggregation, and each layer is allocated with an independent CUDA kernel.
7. The method of claim 6, wherein edge preserving constraints are employed in combination with feature propagation mechanisms to ensure geometric feature consistency between different resolution layers during downsampling.
8. The method of claim 6, wherein statistical moment conservation and distribution matching are combined to ensure consistency of attribute characteristics among different resolution layers during downsampling, and probability distribution of attribute values is ensured to be consistent among different resolution layers through histogram matching.
9. A geospatial grid processing system based on CPU-GPU cooperation, characterized in that the system comprises an input layer, a CPU processing layer, a GPU processing layer and an output layer, and the geospatial grid is generated by applying the method according to any of claims 1-8.
10. An electronic device comprising a processor and a memory, the memory storing machine-executable instructions executable by the processor to implement a method as claimed in any one of claims 1-8.

Description

Geographic space grid multi-resolution streaming construction method based on CPU-GPU cooperation Technical Field The invention relates to the technical field of Geographic Information System (GIS) and high-performance computing intersection, in particular to a method and a system for constructing a geographic space grid multi-resolution streaming based on CPU-GPU cooperation. Background The geospatial grid is used as a core carrier for integrating multisource geospatial information and realizing standardized management and efficient utilization of spatial data, and is an essential basic data model in the fields of digital earth, smart cities, geological exploration and the like. The construction quality and efficiency of the geospatial grid directly influence the effects of a series of applications such as subsequent spatial analysis, visual presentation, decision support and the like. In the development process of the geospatial grid, the core value of the geospatial grid is always developed around 'realizing ordered organization and high-efficiency application of large-scale space data', and along with the rapid development of high-precision geospatial data acquisition technologies such as oblique photography, laser radar and the like, the geospatial data volume is increased in an explosive manner, which brings unprecedented challenges to efficient construction of the geospatial grid, namely how to rapidly generate the geospatial grid with precision and suitability based on massive multi-source data, and especially to construct a multi-resolution geospatial grid pyramid supporting efficient visualization and analysis, so that the geospatial grid pyramid becomes a core technical bottleneck in the application field of the geospatial grid. From the development context of geospatial grid processing technology, early researchers relied on CPU serial computing methods to complete the construction and simplification of geospatial grids. In the mode, generation of each unit of the geospatial grid and construction of topological relation are sequentially carried out, so that the generation period of the geospatial grid is extremely long, and the quick construction requirement of the geospatial grid in a massive data scene can not be met completely. With the rise of GPU parallel computing technology, partial researchers start to attempt to introduce GPU computing power into the geospatial grid construction process, and two main flow schemes of the current geospatial grid processing field, namely a pure CPU processing scheme and a partial GPU acceleration scheme, are formed. The pure CPU processing scheme is used for completing all construction links of the geospatial grid in a whole process by depending on the CPU, although the control logic of the geospatial grid is relatively simple and easy to debug, the construction efficiency of the geospatial grid is low due to the limitation of the calculation performance of the CPU, the generation requirement of a large-scale geospatial grid is difficult to deal with, and part of calculation links in the construction of the geospatial grid are transplanted to the GPU by the GPU acceleration scheme, so that the division degree of the CPU and the GPU is insufficient, the heterogeneous calculation cooperative mechanism with reasonable design cannot be designed for the whole process of the construction of the geospatial grid, the whole efficiency of the construction of the geospatial grid is limited, and the core advantages of the heterogeneous calculation of the CPU-GPU cannot be fully exerted. Although the existing method makes a certain progress in the geospatial grid processing, in the presence of large-scale, high-precision and low-delay construction requirements of the geospatial grid, a remarkable short board still exists, and the specific problems are as follows: (1) The existing method adopts a fixed-size blocking strategy in the construction process of the geospatial grid, the non-uniformity of geospatial data distribution is not fully considered, the construction quality of the geospatial grid is directly related to the reasonability of the blocking, the fixed blocking can cause excessive concentration of the geospatial grid units in a part of data-intensive areas, the blocking of a part of sparse areas causes waste of calculation resources, and finally the uneven calculation load of the construction of the geospatial grid is caused, so that the overall construction efficiency is influenced. (2) The traditional method has the defects that when the geospatial grid is constructed, high-resolution geospatial data of the whole target area is required to be completely loaded into a memory, and then operations such as unit generation, topology construction and the like of the geospatial grid are carried out. (3) In the existing method, the basic construction of the geospatial grid and the generation of the multi-resolution geospatial grid are serial processes which are se