Efficient distributed matrix for resolving computational intensity in remote sensing

Remote sensing analysis is a dominant yet time-consuming part of geospatial applications. The performance can be optimized based on distributed computing, but current systems still face significant challenges. Firstly, the spatial characteristics of remote sensing data lead to an uneven distribution of computational intensity (CIT), which characterizes computing loads, including computation and IO, in different spatial domains. Secondly, it is hard to achieve load-balancing without introducing new computational costs, thus increasing the CIT and reducing the overall performance. Therefore, resolving CIT by decreasing and balancing it is an important research issue for distributed remote sensing computing. This paper proposes LBM-RS, an efficient distributed framework based on load-balancing matrix computing for remote sensing. It implements remote sensing applications based on the distributed matrix, representing the algorithms with a matrix computation and constructing multi-dimensional spatial domains to model computational costs for matrix operation tasks. It resolves the CIT with the minimum computation load and dynamic spatial domain decomposition strategy to support global load balancing. We also exploit the IO efficiency from the task staging strategy and the cache-aware memory structure for remote sensing data. In this way, it can reduce the bandwidth burden and memory access frequency, thus decreasing the overall CIT. Finally, we evaluate the proposed approach on both real and synthetic datasets, and the results demonstrate significant advantages in computation and communication efficiency compared to the benchmarks.