CUDA-based parallelization of time-weighted dynamic time warping algorithm for time series analysis of remote sensing data

Highlights

• CPU-GPU is used to solve the problem of computing intensive remote sensing images.

• The multithreading architecture of GPU is used to process independent pixels in parallel.

• Using the multi-level cache characteristics of CUDA to improve the memory access efficiency of the algorithm.

Abstract

The time weighted dynamic time warping (TWDTW) algorithm is an important algorithm for the time series analysis of remote sensing images. However, due to the limitation of computational complexity, it is difficult to apply this algorithm to large datasets. Therefore, combined with the characteristics of the TWDTW algorithm. This paper proposes a time weighted dynamic warping time parallel algorithm based on CUDA. First, a GPU many-core stream processor is used to process independent pixel data in parallel, and the cumulative cost matrix is established through a multithreading architecture to reduce the dependence between data. Then, the memory model is optimized to reduce memory access transactions in order to improve the memory access efficiency of the algorithm. Taking the Sentinel-2 remote sensing image as experimental data, the algorithm based on different time series lengths, data scales and thread organizations is experimentally verified, and the execution performance of the algorithm under different conditions is explored and analyzed. The experimental results show that this method can significantly improve the computational efficiency of the algorithm.

Introduction

In recent years, changes in human activities have intensified changes in the global surface (Battude et al., 2016). A time series analysis of remote sensing images can accurately and efficiently identify these changes and has become an important technical means to obtain land use information (Gomez et al., 2016). The sentinel series satellites of the European Space Agency (ESA) and the spot series satellites of the French Center for Space Research (CNEs) have been launched one after another, providing researchers with satellite image time series with high time and spatial resolutions (Pelletier et al., 2016). The rapid development of remote sensing earth observation technology not only provides massive data for time series analysis, but also brings the problems of a large amount of file data and a long processing time for its application, which makes it difficult to meet the increasing real-time demand of GIS spatial analysis (Zhu et al., 2018). Therefore, how to improve the computational efficiency and reduce the time consumption of the time series analysis has become a prominent research area for current scholars (Coelho et al., 2017).

Among the algorithms for time series analysis of remote sensing images, the time weighted dynamic time warping algorithm (TWDTW) is used for time series pattern matching of remote sensing images (Maus et al., 2016). Research shows that it has achieved remarkable results in crop classification (Gella et al., 2021), seasonal information extraction (Narin et al., 2021), vegetation type identification (Cheng and Wang, 2019), land use, and land cover mapping (Viana et al., 2019). The TWDTW algorithm is improved from the dynamic time warping (DTW) algorithm. The working principle of DTW is to compare the similarity between known event patterns and unknown time series (Sakoe and Chiba, 1990). Due to its computational characteristics, the algorithm has a time and space complexity of $O\left(n^2\right)$, which limits its application in large datasets (Oliveira et al., 2018). Based on the above considerations, many scholars are committed to the methods and technologies of high-performance parallel optimization of the algorithm. With the maturity of universal programming tools, such as unified computing device architecture (CUDA), researchers have performed much research on the parallelization of the algorithm on multicore computing platforms, achieving remarkable results (Aldinucci et al., 2021). Zhu et al. (2018) proposed an algorithm to find the possible starting position of similar local subsequences. This method checks the local optimal combination for matching, but the remote sensing data are arranged neatly and cannot benefit from the method of searching irregular subsequences. Xiao et al. (2014) proposed a DTW parallel algorithm based on prefix calculation. The algorithm uses a specific data dependency transformation to solve the problem of instance size limitation. However, the algorithm has a high query cost and a poor scalability of low-dimensional data. Zhang et al. (2012) introduced KNN to estimate the lower bound and allocated the path search and lower bound estimation to two cores for calculation. Each calculation requires the transmission cost matrix, which has become a performance bottleneck. The above literature provides an idea for optimizing the DTW algorithm through GPU parallel computing, but it is not completely suitable for remote sensing image processing. A spatial parallel TWDTW (SP-TWDTW) is proposed. Oliveira et al., (2018) considered the first law of geography and determined the types of features in the study area by coordinating the available core analysis timeline and spatial autocorrelation technology, the experimental results show that the response speed is increased by 11 times. In this scheme, the fine-grained parallel strategy is selected, and the parallel design is carried out at the operation level of the cost matrix. Each thread is responsible for the operation of a diagonal unit without dependency. However, this scheme does not make full use of the advantages of the GPU programming model (Wu et al., 2017). The calculation of the cumulative cost matrix requires frequent access and access in the video memory. Each access requires hundreds of clock cycles, which greatly increases the access overhead. Different diagonal lengths will lead to uneven task allocation among threads, which will cause some threads to be idle, resulting in a waste of GPU resources. The important feature of remote sensing data is the long revisit period and wide spatial range of satellites, which leads to an imbalance of the spatial axis and time axis of remote sensing time series datasets. This is very different from other time series data applications in the original target algorithm. The abnormal memory distribution limits the memory throughput and instruction throughput, which seriously affects resource utilization. There are currently few parallel methods for analyzing time series based on the characteristics of remote sensing images. Therefore, this paper proposes a time weighted dynamic time warping parallel algorithm based on CUDA. Combined with the geospatial data abstraction library (GDAL), the data structure is reorganized according to the memory access principle in the writing link to design a multithreaded architecture and memory access model. Then, the cumulative cost matrix is established by using multithreading in parallel, which is separated from the pattern matching stage to achieve partial fine-grained parallelism. Then, the multilevel cache of the GPU is used to compare the subsequence with different modes in different cores, and the final cumulative cost array is returned to the CPU, which is good at logic control and responsible for classifying and outputting labels. Finally, the computational efficiency of parallel algorithms under different conditions is compared and analyzed, and the effects of thread organization, time series length and number of patterns on the performance of the algorithm are discussed to provide a reference for the time series analysis and optimization on heterogeneous platforms.