| Hyperspectral image(HSI)is a three-dimensional cubic data that provides both spatial and spectral information.Since hyperspectral data contains a large amount of spectral reflectance data collected from hundreds of continuous bands,it makes it possible to effectively distinguish the subtle differences between different materials.As a result,HSI is used by researchers in a wide variety of applications,such as meteorological monitoring,agricultural development,food safety,military reconnaissance,medical diagnosis and so on.Among them,HSI classification task is a fundamental analytical task in many fields of remote sensing,so it has important research significance for the field of hyperspectral remote sensing.With the development of artificial intelligence and neural networks,deep learning has become the mainstream research direction for HSI classification.However,the current algorithms for extracting spatial-spectral features using deep learning still have some shortcomings,specifically including the following aspects: 1)the extracted spatial features and spectral features lack guidance;2)most of the designed networks are not strong for global feature extraction;3)the information source of the features is relatively single.To overcome the challenges posed by the above problems,three deep learning-based spatial-spectral feature classification algorithms are proposed in this paper.In order to address the issue of lack of insufficient guidance of feature maps in the network,this paper designs a HSI classification algorithm based on a feature guiding fusion network of spatial-spectral information.The algorithm first puts HSI through the spatial mapping module and the spatial-spectral mapping module respectively to extract the respective corresponding feature maps.After that,the two feature maps are successively put into the multi-guiding block and the self-guiding block for guidance to obtain richer features.The feature maps after guiding continue to be fused by bilinear pooling,and the resulting feature maps are used for the final classification.The effectiveness of the algorithm for HSI is verified in experiments on three hyperspectral datasets.Most of the existing deep learning classification algorithms mainly use the convolutional neural network as the backbone of the network,which leads to the inability of the network to extract the global features of the data.Therefore,this paper proposes a multilayer perceptron(MLP)-based HSI classification algorithm.The algorithm designs an improved Vision Permutator that is composed of MLP structure,which enables it to capture global features of HSI.In addition,by using the designed hierarchical fusion strategy,different dimensional features in HSI can have more comprehensive interactions.The experimental results show that this hierarchical fusion strategy is effective,and the algorithm is able to achieve competitive classification results.Considering that using only the spatial and spectral information of the HSI itself may still have the problem of poor feature recognition.In this paper,we introduce coordinate information and propose a deep learning-based HSI classification algorithm based on MLP in combination with the designed Involution MLP(InvoMLP)structure.On the one hand,the InvoMLP module is built based on InvoMLP for mining the spatial-spectral information of HSI.On the other hand,the network also proposes a coordinate information extraction module,which not only enriches the diversity of information but also further complements the global spatial features of the feature map.By fusing spatial-spectral information with coordinate information,more robust features can be obtained.The usefulness of coordinate information for classification is experimentally verified,and after comparing it with other algorithms,it is demonstrated that the algorithm can achieve better classification performance. |
