Research On Hyperspectral Image Clustering Algorithm Based On Anchor Graph

Hyperspectral remote sensing images,which have high spectral resolution and can fully reflect the fine features of geological features,have become popular in fields such as agriculture,military and scientific research,thanks to advancements in remote sensing imaging technology.Hyperspectral images（HSIs）contain a wealth of spatial and spectral information,with hundreds or even thousands of continuous spectral bands and a high correlation between the bands.The fine classification and identification of hyperspectral images based on the differences in spectral characteristics between different features has received a lot of attention.Clustering analysis is a powerful data analysis method that aids feature monitoring,surveying and delineation by clustering hyperspectral images.Clustering is a traditional unsupervised learning method that classifies similar unlabeled hyperspectral image data into the same class.K-means and fuzzy clustering methods are widely used in hyperspectral image clustering due to their low computational complexity.When the distribution structure of hyperspectral data is non-convex,these algorithms will fall into local optimum,making it difficult to achieve great clustering performance.And the graph theory-based clustering algorithm has become a research hotspot for researchers at home and abroad because it is not restricted by the type of data distribution and can cluster different types of data.The graph theory-based clustering algorithm must construct a graph that fits the data distribution,with each data point serving as a node and the graph’s edges representing the similarity measure between two data points.Existing graph theory-based clustering algorithms can produce good clustering results,but with the dramatic increase in the volume of HSI data,it is critical to obtain clustering results with high accuracy and low computational complexity for the practical application of hyperspectral images.Based on this,two algorithms are proposed under the traditional graph theory-based clustering algorithm and fuzzy clustering algorithm,respectively,in this thesis as follows:（1）To address the problem of high computational complexity of traditional graph theorybased clustering algorithms,the Fast Hyperspectral Clustering Algorithm（FHC-BTA）based on binomial tree anchor points is proposed.Combining the spatial and spectral properties of hyperspectral images,the FHC-BTA algorithm first selects some data points from the original hyperspectral data using the binary tree method as anchor points.Then,a kernel-free similarity map based on the anchor points is constructed to avoid artificial adjustment of Gaussian kernel parameters.Finally,a spectral clustering analysis is performed and applied to the hyperspectral images to obtain the clustering results.Comparing the FHC-BTA method to the traditional clustering algorithm,the experimental findings reveal that the FHC-BTA algorithm achieves superior clustering accuracy in less time.（2）For the problem of low accuracy of traditional fuzzy clustering algorithm in hyperspectral image classification,the hyperspectral fuzzy embedding clustering algorithm based on bipartite graph（FECBG）is proposed.To construct a unified fuzzy clustering model,the FECBG algorithm combines the non-negative regularization term based on bipartite graph with fuzzy clustering model.First,a bipartite graph matrix is constructed to describe the connectivity between the hyperspectral data and the anchor points.The spectral embedding data is then low-dimensionally represented using fast spectral embedding algorithm to reduce the computational complexity of the algorithm.Finally,the non-negative regularization term based on the bipartite graph was introduced to the fuzzy clustering to constrain the solution space of the fuzzy membership matrix,thus reducing the sensitivity of the fuzzy clustering to the initial cluster centers and improving good clustering performance.