Hyperspectral Image Classification Based On Dynamic Group Convolution And Confident Learning

Hyperspectral imaging has been widely used in many real-world applications.Hyperspectral image classification has gradually become a challenging task due to high inter-class similarity,high intra-class variability,overlapping and nested regions.In recent years,the classification of hyperspectral images based on deep learning methods has attracted more and more attention from scholars.Among them,convolutional neural network is one of the most commonly used visual data processing methods.Aiming at the shortcomings of the existing group convolution,this paper proposes a hyperspectral image classification model(DGDN)based on dynamic group convolution.This structure can dynamically and adaptively select the input channels to be connected in each group for a single sample,that is,each group is equipped with a feature selector,so that it can automatically select the most important input channels according to the input image,so as to capture rich and complementary visual features.In this paper,a series of experiments are carried out on four widely used hyperspectral datasets,and the results show that the proposed model DGDN is better than other comparison algorithms.In addition,in real hyperspectral datasets,there are often some mislabeled noise data.In this paper,a data centric confident learning method is introduced to experiment with noisy datasets.It is a method of characterizing,identifying and learning noise labels in a dataset.It is mainly used to improve the training of noise labels by estimating the joint distribution between the noise labels and the real labels,and to identify the noise labels in the dataset,so as to provide clean data for training and improve the classification accuracy of the model.The experimental results show that the confident learning does have an improvement effect in the hyperspectral image dataset with noise.