Classification Of Hyperspectral Image Based On Spatial Pooling Graph Convolutional Network And Federated Learning

Remote sensing hyperspectral image classification is part of remote sensing image interpretation.Remote sensing image interpretation has emerged in the last half century with the booming development of remote sensing satellites and other long-range sensors,and its core lies in analyzing image information to guide subsequent tasks.With the explosion of deep learning technology in recent years,remote sensing image interpretation has benefited from it and made leaps and bounds.In particular,for remote sensing hyperspectral image classification,the technology in this field has transitioned from the early manual interpretation and traditional algorithms to today’s deep learning led by convolutional neural networks.Convolutional neural networks have greatly improved the accuracy of classification with their unprecedented ability to fuse complex hyperspectral neighborhood information and abstract highly semantic features.On the other hand,numerous representative models have emerged in the field of deep learning in recent years,and these models have achieved good results in various fields.Therefore,their contributions to hyperspectral image classification also deserve a closer look,The graph convolution,spectral clustering,and federation learning used in this paper are taken from these representative models,and the model architectures and experimental methods are redesigned for hyperspectral image classification tasks to demonstrate their effectiveness.The main works are as follows.1)A hyperspectral image classification method based on spatially pooleding graph convolution is proposed that captures the relationships between image element neighbors and naturally fuses these elements on the relationship graph.To achieve this,we construct graph structures over local neighborhoods and extract features using graph convolution networks.A major technical problem that needs to be addressed is that the pooling operation of graph convolution is less regulated than that of convolutional networks.Although many works have been proposed for graph pooling,they are limited by the applied data structure and cannot be adapted to hyperspectral images.Therefore,this work designs a graph pooling operation with spatial pooling as the goal.This pooling operation ranks the importance of image elements by spatial location and retains the image elements with high importance.Experiments show that this spatial pooling graph convolution model can beat the popular convolutional neural network model on a typical hyperspectral classification dataset.2)A federated learning method based on deep spectral clustering is proposed As a classical graph learning model,spectral clustering is well integrated into the wave of deep learning and continues to contribute in the frontier.Compared with K-means clustering,spectral clustering can better capture the intrinsic flow structure of the data.This work takes deep spectral clustering as a starting point,and uses the popular federation learning in recent years as an application direction to design a federation learning architecture based on deep spectral clustering to improve the performance of federation learning.The introduction of deep neural networks enables the method to better extract the features of model updates,and the introduction of spectral clustering enables the method to better capture the global update structure in federal learning.Experimental results show that the method has better performance in capturing heterogeneous data compared to previous federal learning architectures.3)A general federated hyperspectral image classification framework is proposed.With the rise of big data,federated learning tries to learn good models from huge amount of siloed data.As a consistently rising research hotspot in the last five years,the entire machine learning community and large companies are increasingly focusing on the value of federation learning in typical future application scenarios.Fragmented content and siloed storage is becoming the way data exists in more and more domains,and hyperspectral image processing will be no exception.Therefore,this section of work hopes to build this future scenario and present some possible problems and solutions.Overall,the core goal of the research content of this paper is to improve the image classification accuracy and to advance the hyperspectral image classification with reasonable ideas and solutions in other possible development directions.