Based On Prior Information Fusion And Prototype Measurement Hyperspectral Image Classification Algorithm

With the development of hardware technology for hyperspectral sensors,the spatial resolution of acquired image data has become increasingly refined and the number of spectral bands has grown.In order to fully extract feature information,deep learning algorithms have been applied in the field of hyperspectral image classification.However,during the model training process,issues such as noise interference and insufficient subspace utilization can arise,especially when the number of labeled samples is limited in certain scenarios,posing great challenges for supervised learning methods in hyperspectral image classification.To address this,scholars have proposed the concept of a small-sample prototype network.However,during its training phase,problems such as weak feature representation ability and misaligned cross-domain data distribution may occur.Therefore,this thesis proposes a supervised deep learning method for noise matrix removal,and then considers the lack of labeled samples in practical application scenarios by proposing two small-sample measurement network models based on multi-feature and covariance measurements,respectively.The main works are as follows:(1)Aiming at remove sparse noise matrices and improve the utilization rate of subspaces,a residual splitting network for low-rank matrix recovery is proposed.First,an algorithm using non-exact Lagrange multipliers is used to recover a low-rank matrix,which is then sent to a residual classification network with an interactive channel attention mechanism for extracting detailed features,ultimately leading to an end-to-end hyperspectral classification network.Experimental results verify the effectiveness of lowrank matrix recovery in removing noise data from hyperspectral images and improving classification accuracy.(2)Aiming at obtain more accurate feature prototypes and alleviate the problem of misaligned cross-domain data distribution,a multi-feature measurement cross-domain adversarial small-sample classification method is proposed.By using conditional adversarial domain adaptation strategy and multi-feature measurement,both source and target domain learning are performed,which enables the discovery of transferable knowledge and learning discriminative embedding models in the source domain.Experimental results demonstrate that the model’s feature prototype expression ability is enhanced and improve the classifier performance.(3)To achieve the non-linear transformation ability of the network and improve its resistance to data anomalies,a small-sample classification network based on covariance measurement is proposed.The source domain dataset is first reconstructed and preprocessed using morphological operators,followed by designing a lightweight classification network.Finally,the covariance measurement is used to support the distance between the class prototype features of the support set and the query set feature vectors.Experimental results show that this method achieves good performance in smallsample classification,exhibiting outstanding results in both classification accuracy and effect charts.This thesis proposes a method for fusing prior information and constructing a new prototype measurement method to achieve classification of hyperspectral images.These methods allow the model to have clearer decision boundaries and result in better distribution of land cover data mapped to feature space.