Hyperspectral Image Classification Based On Domain Adaption

Hyperspectral image classification is an important content of hyperspectral remote sensing earth observation technology,which has important applications in many fields such as water quality detection,fine agriculture,geological survey and so on.The high-dimensional spectrum and high correlation between spectral bands of hyperspectral images make traditional hyperspectral image classification methods face great challenges.Therefore,the use of deep learning algorithm without manual feature design for effective feature extraction and classification is the mainstream trend in this field.However,due to the differences in imaging equipment of hyperspectral images and changes in captured environment,the trained classification model is difficult to be applied to newly captured hyperspectral images and training a new classifier for each hyperspectral image restricts the popularization and development of large-scale earth observation tasks.Based on the deep learning theory,this paper studies the hyperspectral image classification method based on the unsupervised domain adaptive network,and realizes the cross-data hyperspectral image classification model transformation and information sharing without manual labeling.The main work of this paper is as follows:First of all,in order to solve the problem that the spectral band coverage of different sensors is different and leads to the inconsistency of feature space,an adaptive hyperspectral image classification method based on cross-data set domain is proposed to complete the cross-data set transformation of the classification model.The network introduced cross-data of auxiliary tasks to the relevant task,The domain alignment module is designed under the guidance of the cross-data set of the auxiliary task to minimize the feature difference between the source domain and the target domain.The task assignment module is used to learn the task-related source domains based on the shared domain alignment module’s partial network parameters.The designed domain adaptation module transfers alignment and classification capabilities to the target domain.Thus,the transformation of hyperspectral image cross-data set classification method is realized.Secondly,aiming at spectral feature shift of the same ground object caused by the capture of environmental changes,we proposed an adversarial learning domain adaptation hyperspectral image classification method to realize cross-data set information sharing of classification model.The whole network is trained with adversarial learning.With the source domain initials generator and discriminator,adjust the discriminator on both domains and fine-tune the generator on the target domain in an adversarial manner by multiple times.The training process of adversarial learning drives the features of the target domain into the feature space of the corresponding class of the source domain,aligns the source domain and the target domain while maintaining different class boundaries,so as to realize the unsupervised domain adaptation.In addition,the network uses a generator based on a variational auto encoder to learn the distribution of inputs and generate a more general representation,which is more robust to classification problems across datasets.To sum up,for the difficulties of hyperspectral images classification model in the practical application,this paper puts forward to use the deep learning algorithm and the knowledge of cross-data set to reduce the reliance on the number of target domain samples and realize transformation of classification model.This paper also puts forward to an adversarial learning method to realize information sharing and improve the practicability of the classification model.These methods provide a reference for the practical application of hyperspectral image classification in large-scale earth observation tasks.