Non-Gradient Optimized Graph Convolution Neural Network And Its Application To Hyperspectral Image Analysis

With the rapid development of computer and artificial intelligence technology,neural networks have been widely applied in various fields,including speech,text,and image processing.However,the vast majority of real-world scenarios,such as social networks,have unambiguous topological structures,which makes it difficult for traditional neural network models to be fully competent.In recent years,significant progress has been made in the research of structural neural networks.Graph convolutional neural networks(GCNs),as the most representative achievement of geometric deep learning methods,have provided powerful support for graph data processing,as well as new ideas for solving traditional learning problems such as hyperspectral image classification.However,the current graph convolutional neural networks face two key problems in the optimization and application process:(1)in the optimization process,the graph convolutional neural networks based on gradient descent not only easy to fall into the local optimum but also may cause the problem of gradient disappearance or gradient explosion.Therefore,it is of clear practical significance to study the non-gradient-based training method;(2)in the application process,such hyperspectral images often have the problems of large data volume and difficult label acquisition.On the one hand,the transductive graph convolutional neural networks lack scalability on large-scale data.On the other hand,unsupervised graph convolutional neural networks have greater room for improvement in classification accuracy.Based on the above background,this thesis aims to explore the effectiveness and feasibility of non-gradient graph convolutional neural networks from both theoretical and application aspects.From the theoretical aspect,this thesis focuses on the simplification and optimization of GCN using a random weighting strategy and evolutionary search strategy.In the application aspect,this thesis attempts to apply the proposed GCN models to two typical classification problems of hyperspectral images,i.e.,supervised classification and unsupervised classification.The main contributions of this thesis include:(1)The classic randomized neural networks cannot model graph structures,leading to poor robustness.To overcome this problem,a general semi-supervised learning framework based on randomized GCN is proposed.Firstly,the concept of random graph convolution is proposed,and it is proved that such random generated graph convolution operation can not only ensure excellent discriminative ability but also have good diversity.Secondly,the classical ELM and RVFL networks are extended to the non-Euclidean domain by using random graph convolution,to make them the ability to process graph data.Finally,based on the diversity of random graph convolution,an ensemble-based random GCN method is proposed,which effectively improves the robustness and stability of the model.Extensive experiments show that the proposed framework not only retains the advantages of the classical random neural network but also can obtain better classification performance.(2)The previous randomized GCN cannot guarantee the global optimal graph embedding,which will lead to a large variance and requires more neurons.To address the shortcoming,an evolution-driven hybrid randomized GCN is proposed.In this method,the parameters of the random graph convolutional layer are regarded as the population of the evolutionary method.The fitness function is defined as the classification error.To search the global optimal graph embedding,an improved adaptive differential evolution algorithm(JADE)is adopted.In addition,to improve the diversity of population behavior in the evolutionary algorithm,a novelty search strategy is introduced into JADE,which effectively improves search efficiency.The presented hybrid approach is a general framework and is suitable for any metaheuristic algorithm.Experimental results show that the proposed algorithm has excellent robustness and effectiveness on both graph and non-graph datasets.(3)Although the proposed randomized GCN models simplify the learning of GCN models,they cannot be applied to large-scale hyperspectral image classification tasks due to their transductive learning mechanism.To overcome this issue,this thesis extends the randomized GCN using inductive learning.Firstly,the pixel-wise classification of hyperspectral images is transformed into a graph-level classification problem by constructing local neighbor graphs.Secondly,a global graph pooling strategy is used to generate graph-level representations in an inductive manner,which makes it able to process data with min-batch.Finally,the model is trained efficiently by solving the global closed-form solution,which significantly reduces the training cost.Experimental results on several hyperspectral images show that the proposed method has good scalability and can achieve competitive classification results with higher computational efficiency.(4)Hyperspectral image classification often suffers from limited training sample problems.To enable GCN to classify hyperspectral images without labeled samples,this thesis proposes a novel hypergraph convolutional subspace clustering model.Firstly,the hyperspectral image is converted to a hypergraph representation,which effectively encodes the high-order structure information of the image.Secondly,by generalizing the traditional subspace clustering into the non-Euclidean space,a robust hypergraph convolution self-expression model is proposed.Finally,a simple and efficient multi-hop aggregation mechanism is proposed to learn the long-range dependencies between graph nodes.A large number of experimental results show that the proposed method is significantly better than many hyperspectral image clustering methods.Through the above research at the theoretical and application levels,a series of new algorithms for non-gradient optimized graph convolutional neural networks are proposed,which not only effectively improve the efficiency and performance of graph convolutional neural networks,but also provide effective new ideas for the processing of hyperspectral images.