Известия Томского политехнического университета: Инжиниринг георесурсов (Sep 2017)
Using different computing systems to solve the automatic cloud classification problem according to MODIS satellite data by probabilistic neural network
Abstract
The relevance of the research is caused by the necessity to develop algorithms and software to classify the cloud types based on single-layer cloud on the satellite images received from MODIS spectral radiometer used in Terra and Aqua remote sensing Earth satellites with the usage of high-performance systems. The main aim of the study: effective and fast analysis of 5416-8120 single-layer cloud full scale satellite images received from MODIS spectral radiometer with the help of the probabilistic neural network detecting 27 cloud types. The methods used in the study. To carry out the task the authors used the methods of paralleling the processing, neurocomputing, computer vision and texture analysis algorithms, classification algorithms, technologies of high-performance processing for multi-core shared memory systems (OpenMP), graphics processing units (CUDA) and distributed systems (MPI). The results. The classifying procedure based on probabilistic neural model compares all the fragments from the given image with the patterns from the training set classified by experts. It needs to compare texture features of each fragment with features of some thousands patterns and therefore it leads to significant time costs. The algorithm allows splitting the given input into a set of small images that can be processed independently by some computational devices and devices supporting the processing of simultaneous tasks. The paper compares the performance of three approaches for parallel processing that are multi-thread computation based on multi-core central processing units (CPUs), multi-thread computation based on graphics processing units (GPUs) and distributed processing implemented by computational cluster. The latter uses worksharing between different processes with independent address spaces and the approach includes two methods for speed-up the processing based on data distribution and task sharing. Each approach was described in detail and its performance was estimated by analysis of MODIS' full scale image. It's shown that the usage of distributed processing or/and multi-thread GPU computation for performance of single-layer cloud classification task based on probabilistic neural model has significant performance advantages not only in comparison with the classic sequential algorithm but also with its multi-thread version for many-core CPUs.
