Hyperspectral Remote Sensing Image Classification Based On Fuzzy Spatial-spectral Features

Hyperspectral image contains rich spectral information and spatial distribution information of ground objects,and has a wide range of applications in agriculture,ecology,marine,military and so on.The task of classification of hyperspectral remote sensing images has been receiving much attention in the field of remote sensing.Many deep learning methods have achieved some results here,but there are still some limitations.Due to the inherent high dimension of hyperspectral image data,lack of label samples and band redundancy,as well as the adverse influence of mixed pixels,it is easy to be polluted by noise points,leading to problems such as greater intraclass difference,stronger interclass coupling,and easy misclassification of boundary pixels,which seriously affect the classification performance and bring certain difficulties to the classification task.The deep learning model based on spatial-spectral joint features fully exploits the hidden features of data by making exhaustive use of spatial information and spectral information of hyperspectral remote sensing images,solves the spectral heterogeneity and model degradation problems in hyperspectral remote sensing image classification tasks,and this has led to significant improvements in the performance of the classification method.The fuzzy learning mapping feature is used to overcome the interference of neighborhood pixels,weaken the noise pollution,enhance the representation capacity of the network,further stand out the significant information of the target pixels.Thus,based on the traits of "unity of spectral and images" in hyperspectral remote sensing data,in this paper,we investigate hyperspectral remote sensing image classification methods based on fuzzy spatial-spectral features,contraposing the intricate relationship and nondeterminacy among pixels in hyperspectral remote sensing image classification task,combining fuzzy learning with attention mechanisms,in order to preferably depict the correlation between pixels.To settle the matter of boundary pixel misclassification and noise pollution,thus significantly enhancing the classification performance of hyperspectral remote sensing images.The dissertation includes the following contributions:(1)In view of the uncertainty of hyperspectral remote sensing data and problems with misclassification of boundary pixels,insufficient use of spatial information and degradation of the fit in the process of classifying deep learning models,proposed a network based on fuzzy spatial-spectral features for the classification of hyperspectral remotely sensed images.Firstly,an asymmetric convolution fuzzy module is designed to enhance the representation ability of deep learning convolutional layer,highlight the target pixel information,and capture more distinguishing and discriminant features.Secondly,in the process of spectral feature extraction,gated recurrent unit was used to capture the relevant information in long sequences to filter out irrelevant information,reduce network parameters,and improve the efficiency of model training.Three hyperspectral remote sensing data sets were selected in the experiment to verify the effectiveness of the algorithm.As the extracted the spatialspectral joint features are more distinguishable and discriminant,the experimental results manifest that the features extracted from the proposed network based on fuzzy spatialspectral features are more related to the target features,which not only solves the problem of the boundary pixel error,but also effectively deals with the uncertainty of hyperspectral remote sensing data.Contrasting the subsistent hyperspectral remote sensing image classification models,more accurate classification accuracy can be obtained with the proposed model.(2)Aiming at the problems of hyperspectral remote sensing data’s high redundancy,insufficient labeling samples and insufficient discriminant feature extraction by existing methods,putting forward a semi-supervised network model based on fuzzy attention spatialspectral joint features for hyperspectral remote sensing image classification from the perspective of both spatial and spectral features.Through end-to-end learning strategy,this method introduces fuzzy learning to overcome the distraction of noise to improve the signalto-noise ratio and decrease the affect of noise on the uncertainty of classification.In addition,the channel attention mechanism is combined to acquire the correlation between channels,enhance the expression ability of network features,highlight important features and suppress non-relevant features,and capture the more discriminant spatial-spectral features.Through experimental testing and comparative analysis on four hyperspectral remote sensing data sets.These results demonstrate that,under the condition of limited labelled samples,the proposed method effectively uses the obtained spatial-spectral joint features,achieves preferably experimental accuracy,and enhances the classification manifestation of the model.