Research And Application Of Small Sample Hyperspectral Image Ground Object Classification Algorithm Combined With Spatial-Spectral Information

Hyperspectral images have tens or even hundreds of continuous and subdivided wavebands.It can image the target area more accurately and carefully,and it has a lot of rich spectral and spatial information that is useful for the classification of ground objects.Therefore,hyperspectral images are widely studied and applied in precision agriculture,atmospheric monitoring,mineral exploration and other fields.With the continuous development of hyperspectral remote sensing imaging technology,hyperspectral images have become easier to acquire,but the acquisition of marker samples is time-consuming,laborious and very difficult.However,the overall features of hyperspectral images are difficult to obtain from small sample of marker samples,therefore,small sample hyperspectral image classification has become the focus of research in remote sensing.Most of the small sample hyperspectral image classification methods ignore the spatial-spectral information and global spatial correlation of the features,and only extract the depth features of the small sample labeled data,and fail to make full use of the large amount of unlabeled sample information.To address the above problems,the main work of this thesis is as follows:（1）A spatial-spectral generative adversarial network with active learning strategy is proposed for small sample hyperspectral image ground object classification model,in which the spatial-spectral generative adversarial network（AEGAN）consists of two generators（G₁,G₂）and a discriminator（D）.Among them,the generator G₁ generates one-dimensional spectral bands,and the generator G₂ generates three-dimensional data blocks,while adding cross-entropy loss to the encoder part within the G₁ and G₂ structures to assist in supervised sample generation,so that the generated samples are closer to the real data.The discriminator D extracts the spectral and spatial features of the samples and obtains the joint spatial spectral information by means of feature concatenation.Under the condition of small sample hyperspectral images,an active learning strategy considering spatial neighborhood is introduced.This strategy not only considers the spectral information,but also makes full use of the location correlation of spatial neighborhood,and its iterative selection of information-rich samples and accurate manual labeling,constant updating of the training set and fine-tuning of the classification performance of the AEGAN network,achieve the integration of the AEGAN network with active learning strategies.（2）A spatial-spectral features and similarity metrics for small sample hyperspectral image ground object classification model（SS-CNet）is proposed,which consists of a spatial-spectral siamese network,a similarity network and a clustering supervision task,the training task is achieved using weighted values of multiple loss functions combined with back propagation.The spatial-spectral siamese network is used to extract spectral and spatial information and obtain joint spatial-spectral features.Under the condition of small sample hyperspectral images,the training method of meta-learning is used to improve the generalization ability of the model by randomly selecting several training tasks consisting of support and query sets in the training set.The similarity network calculates the similarity score between samples and obtains the classification results based on the similarity score.The clustering supervision task is introduced,and the clustering labels obtained by K-means are assigned to unlabeled samples as supervision information,which can extract more discriminative information from a large number of unlabeled samples in addition to making full use of the depth features of small sample data,thus improving the classification effect of the model.（3）A global spatial-spectral graph attention network is proposed for small sample hyperspectral image ground object classification model（SSCG）,which consists of a spatial-spectral siamese network and a global graph attention network.The spatial-spectral siamese network is used to extract joint spatial-spectral features.The number of graph nodes and computational effort are greatly reduced by introducing superpixel segmentation,which makes each superpixel act as a graph node by pixel to superpixel conversion.The graph attention network is combined with channel attention and position attention mechanisms under small sample hyperspectral image conditions,focusing on the importance of channel and position,and using graph attention coefficients so that the correlation between global node features is better incorporated into the model.The fusion of the local features extracted by the spatial-spectral siamese network and the global features extracted by the global graph attention network leads to better classification results even under the conditions of small sample hyperspectral images.（4）Design a hyperspectral satellite image interpretation system,which mainly consists of data pre-processing module,image interpretation module and image drawing module.The data pre-processing module of hyperspectral images includes operations such as data downscaling,3D data block interception and background removal,and the processed data can be used to realize the classification of hyperspectral images.The image interpretation module is based on four hyperspectral image classification algorithms to realize the image interpretation function of the system.The image drawing module includes the drawing of 3D stereo map and spectral curve map.This thesis provides users with a multi-functional multi-algorithm hyperspectral satellite image interpretation system.