Dual-Level Convolutional Neural Network For HSI Classification

Hyperspectral images provide rich spectral information,which can identify more objects that are difficult to recognize in other remote sensing images.Therefore,hyperspectral imaging is widely used in agriculture,geology,oceanography,environmental science,military and disaster prevention and control,etc.The research on hyperspectral image classification methods is also increasing.The high dimensionality of hyperspectral images provides sufficient information,but also causes the difficulty of classification and increases the computational burden.It is very important to develop effective and robust methods for hyperspectral image classification.At present,deep learning methods are used in many fields.Especially convolutional neural network is suitable for image classification task.In this thesis,a new hyperspectral image classification model,the dual-lavel convolutional neural network,is proposed.The expansion of the network is realized from two aspects of width and depth.The dual-level convolutional neural network adopts the parallel multi-branch structure and serial structure to extract features of different scales and levels.First of all,the network uses the Inception module,which extracts the features of different scales by the parallel and multi-branch structure.At the same time,the residual connections are added between the Inception modules.It reduces the possibility of gradient disappearance,which is helpful to improve the expression ability of the network.Secondly,this thesis uses the method of dual-level serial connection to build the network,which has a low-level feature extraction layer and a high-level feature extraction layer,making full use of the strong complementary and related information between different hierarchical layers.Then,using the fusion strategy to fuse the features extracted from the two layers,so that the final fully connected layer can obtain more comprehensive image information.Finally,the pixels to be classified with their neighboring pixels are sent to the convolutional neural network as the input to realize the spectral-spatial joint classification of hyperspectral images,so as to obtain a better classification effect.Experiments on three real hyperspectral datasets show that the dual-level convolutional neural network proposed in this thesis performs better than other comparison methods in terms of visual effect and evaluation indicators.The architecture of the network is universal and has good generalization performance,so it is hardly to occur the over-fitting phenomenon.