Hyperspectral Image Classification Based On Spectral Partition And Spatial-spectral Attention Mechanism

Hyperspectral image classification aims to assign all pixels in a hyperspectral image to a set of specific categories.It is one of the most active research topics in the field of hyperspectral image processing.Different from RGB images,the higher spectral dimension,higher spatial variability and spectral variability of hyperspectral images bring challenges to the task of hyperspectral image classification.Due to the excellent results on natural images processing,convolutional neural network is naturally introduced into hyperspectral image classification.Therefore,this thesis designs a hyperspectral image classification framework based on convolutional neural networks.Inspired by the visual attention mechanism,we design attention modules for the network that improve the performance of the model.Also,we improve the inference procedure of the model to eliminate redundant calculations and improve the efficiency.Specifically,1.In view of the high spectral dimension of hyperspectral images and the existence of the mixed pixels.A hyperspectral image classification framework based on spectral partition strategy and category constraints is proposed.The spectral partition strategy can effectively reduce the dimensionality of the input data,without losing any original spectral information.The category constraint structure groups and constrains the output features of the network,so that the feature extraction of each category is separated and independent from each other,which is more in line with the data characteristics of hyperspectral images.Experiments show that the framework we proposed has better classification performance than existing classification frameworks.2.Aiming at the high spatial variability and spectral variability of hyperspectral images,a spatial attention module,a spectral attention module,and a spatial-spectral attention module are designed.The spatial attention module is used to process the spatial variability of hyperspectral images by extracting pixels that are conducive to classification in the input and suppressing the pixels extracted by interference features;by extracting useful bands and suppressing noise bands,the spectral attention module is used to process the spectral variability of hyperspectral images;the spatial-spectral attention module is designed in conjunction with the spatial attention module and the spectral attention module.Experiments show that the attention module proposed in this thesis can further improve the classification performance of the model,and the modules can be integrated into any CNN-based classification frameworks,which shows a certain degree of scalability.3.Aiming at the problem that the proposed hyperspectral image classification framework based on the spectral partition strategy is relatively complicated in implementation,and has a low training and testing efficiency.Grouped convolution is used to unify the spectral partition and feature extraction process.Thus,an end-to-end model is equivalently designed.And by adjusting the data processing behavior of each layer in the inference stage of the CNN,without changing the parameters of the trained network,the whole image can be fed into the model and performed a one-time inference,thereby avoiding redundancy calculation.Experiments show that the performance of the equivalently designed model can approximate the best results of the original model,and greatly improves the model inference efficiency.