Space-spectrum Joint Classification Of Hyperspectral Images Based On Multi-GPU Parallelism

In recent years,hyperspectral image processing has received wide attention.And it has been applied in many fields of earth science.Hyperspectral image classification is an important research direction of hyperspectral remote sensing.Most of the existing hyperspectral image classification algorithms are based on the combination of spatial and spectral information.These classification algorithms have extremely high computational complexity and cannot meet the requirements of real-time processing.With the rapid development of Graphics-related technologies,the computing capability of GPU has been continuously improved.Using GPU to improve the computational efficiency of the spatial-spectral classification algorithms has become an effective method.However,due to the limitation of memory and computing resources,the size of data that a single-GPU platform can handle is still limited,and the processing efficiency can also be optimized.Therefore,it is necessary to use multi-GPU platform to achieve efficient processing of larger data.The spatial-spectral classification algorithms such as the hyperspectral image classification based on MRF-based spatial prior,and the hyperspectral image classification based on residual neural network have been researched in this thesis.After analyzing the algorithm and the intrinsic characteristics of hyperspectral data,the parallel optimization of hyperspectral image classification algorithm is carried out based on CUDA architecture and multi-GPU platform.The hyperspectral image fast and fine classification system is also developed to realize the quasi-real-time hyperspectral image classification.And the effectiveness and efficiency of the method proposed are verified by real hyperspectral datasets.The main works are as follows:1.We propose a fast algorithm based on multi-GPU platform for hyperspectral image spatial-spectral classification.The hyperspectral image classification algorithm based on MRFbased spatial prior is researched in this thesis.This algorithm models the spectral information by Sparse Multinomial Logistic Regression,and the spatial information is modeled by weighted Markov field.Then it achieves a high classification accuracy.However,due to its high computational complexity and the large data amount of hyperspectral image,the classification algorithm cannot meet the requirements of real-time processing.In order to improve the computational efficiency of the algorithm,we first construct a data flow organization graph for this algorithm based on multi-GPU collaborative interactive,and then design a multi-GPU stream synchronous parallel computing structure.Based on the structure,we implement the multi-GPU parallel optimization of the algorithm on CUDA architecture.The optimization improves the execution efficiency of the algorithm and achieves a high speedup.2.We research a spatial-spectral classification algorithm based on the residual neural network.Then the multi-GPU parallel optimization algorithm is designed based on cuDNN parallel acceleration library.This thesis uses the residual module to extract the spatial and spectral information,and then simplifies the structure of SS-CNN network and adds the batch normalization layer to prevent over-fitting.Then the improved method ResHSI achieves a high classification precision.In order to improve the training efficiency,we use the cuDNN library to build the ResHSI neural network,and then implement the multi-GPU parallel optimization of the training process based on CUDA.The experimental results of real hyperspectral datasets show that the multi-GPU parallel optimization method proposed in this thesis can realize the fast training of neural network and achieve high classification accuracy.3.Based on above algorithms,HSIFC is designed and implemented based on CUDA/MFC framework.This thesis presents the detailed design and implementation of the framework and core modules of the software.We test and apply the system on real hyperspectral datasets such as Pavia University and Pavia Center.