Research On Model Design And Application Of Cube CNN In CPU-GPU Heterogeneous Computing Environment

Hyperspectral image(HSI)classification has been proved significant in remote sensing field.With the rapid growth of spectral information,traditional classification methods have met bottlenecks due to the lack of remote sensing background knowledge and high dimensionality.Deep learning based methods,such as deep convolutional neural network(CNN),can effectively extract high level features from raw data.Meanwhile,it is still a hotspot that how to design a new CNN model which can utilize the spatial and spectral information of hyperspectral image simultaneously.The general purpose graphic processing units(GPUs)have been considered as one of the most common co-processors,and widely used in high performance computing.Recently,to ease the problem of time-consuming training of deep neural networks,more and more researchers are devoted to applying GPUs to deep learning area,which has achieved significant speedup effect.In this paper,we design a Cube CNN model for HSI classification,which takes advantage of CNN model to extract spatial-spectral features for raw HSI dataset.To accelerate the training of Cube CNN model,we propose a GPU-based Cube CNN(GCN)framework for hyperspectral image classification.First,a Parallel Neighbor Pixels Extraction algorithm is designed to enable the framework directly loading raw HSI data into GPU's global memory,and extracting samples into data cube.Then,based on the peculiarity of HSI and cube convolution,we propose a novel Cube CNNto-GPU mapping mechanism that transfers the training of Cube CNN to GPU effectively.Finally,the mini-batch gradient descent(MBGD)algorithm is improved with Multiple Computing United Device Architecture(CUDA)Streams technique,which further speeds up model training in GCN framework.Experimental results show that,compared with state-of-art framework Caffe and Theano,we achieve up to 85%+ and 90+% reduction in model training time without losing accuracy.Experiments across different GPUs demonstrate the good extendibility of GCN framework.