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

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. •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.