Hyperspectral Remote Sensing Image Classification Based On Fuzzy Graph Architecture

Hyperspectral remote sensing images contain rich spectral feature information and a certain degree of spatial information,which have significant practical value for the accurate classification of ground features.Over the last few years,hyperspectral remote sensing image classification has gained increasing attention in the field of remote sensing.However,hyperspectral remote sensing images usually have uncertainty introduced by the problems of “the same object with a different spectrum” and “a foreign object in the same spectrum”caused by atmospheric radiation and noise during the imaging process.Moreover,hyperspectral remote sensing images still have difficulties,such as expensive labeling and dimension disasters.These factors have brought a huge challenge to the accurate classification of hyperspectral remote sensing images.The graph architecture model aims to capture the dependency of samples through the information transfer and aggregation among nodes in the graph and has excellent feature representation capability.However,most state-of-the-art graph architecture models build a graph based on distance measurement,which makes it challenging to fully characterize the complex relationship of hyperspectral remote sensing image data.Fuzzy logic provides an effective similarity measurement method to characterize the complex relationship of uncertainty between hyperspectral remote sensing samples by defining the fuzzy membership relationship between data.Therefore,this paper mainly studies the hyperspectral remote sensing image classification method based on the combination of graph architecture model and fuzzy logic,focusing on the description of uncertainly hyperspectral remote sensing image data in the classification task,and then improving the classification accuracy of hyperspectral remote sensing images.The main research works are summarized as follows:（1）For the noise interference and intraclass and interclass uncertainty of hyperspectral remote sensing image classification,we use a fuzzy similarity measure to replace the traditional Gaussian kernel method to build a graph and introduce an anchor graph structure to reduce the computational cost.Then a spectral clustering algorithm based on a fuzzy similarity measure is proposed.In addition,by iteratively updating the anchors and the fuzzy similarity between the anchors and data points,the algorithm obtains more evenly distributed anchors and a more stable graph.The validity of the algorithm is verified on three hyperspectral remote sensing data sets.Experimental results show that the proposed algorithm achieves better classification results than the state-of-the-art relevant classification algorithms in both the intraclass and boundary regions.（2）By introducing fuzzy logic and fuzzy learning into the graph convolutional neural network,we propose a novel fuzzy graph convolutional network for hyperspectral remote sensing image classification.To overcome the computational complexity of image construction caused by the excessive number of samples of hyperspectral remote sensing images,we adopt the superpixel segmentation algorithm to segment the hyperspectral remote sensing image into a small number of superpixel regions.In addition,we proposed a new graph construction method that obtains the fuzzy similarity between hyperspectral remote sensing data instead of the Euclidean distance metric to characterize the inherent uncertainness.Furthermore,we introduced a fuzzy learning module into the network structure to process the superpixel hyperspectral data and thus reduce the effect of intraclass heterogeneity in hyperspectral remote sensing images on the classification accuracy.Finally,we conducted experiments on three hyperspectral remote sensing datasets to test the proposed method.The experimental results show that the proposed algorithm can achieve class separation and suppression of intraclass heterogeneity better and has higher classification accuracy and better performance.