Data Intrinsic Structure-Driven Deep Learning Method For Remote Sensing Image Classification

Remote sensing image classification which is one of the key tasks in remote sensing image processing has played an irreplaceable role in urban planning,geological exploration,reconnaissance and early warning,and others.In recent years,with the development of deep models,especially convolutional neural networks,breakthroughs of performance in remote sensing image classification have been made.However,the applications of deep models in remote sensing image tasks still face a series of challenges,such as the extensively high spectral dimensions and serious redundancy between different spectral bands,and the limited number of training samples.In addition,the special end-to-end training process in deep learning makes it difficult to utilize the prior knowledge of the remote sensing images,such as the certain spectral curves.This further multiplies the difficulty to apply the deep models in the literature of remote sensing images and leads to the terrible classification performance.Based on the deep models,this paper discovers the potential of the intrinsic characteristics of the remote sensing images in deep learning process and further explores the data intrinsic structure driven deep learning methods to solve the difficulties occurred in the applications of deep learning in remote sensing image classification.The main contents and innovations of the whole paper are listed as follows:Firstly,according to the high spectral dimensions and serious redundancy between different bands as well as the uncertainty of the spectra in remote sensing images,this paper proposes the intrinsic structure-based representation methods for remote sensing images and abstract three intrinsic properties,namely the statistical characteristics,lowdimensional manifold characteristics and generalized clustering characteristics,from the data intrinsic structure.Specifically,based on the imaging characteristics of hyperspectral remote sensing images,the statistical histograms and distribution fitting are used to analyze the statistical distributions of different spectral bands from different classes in hyperspectral data,thus demonstrating the intrinsic statistical properties of the hyperspectral remote sensing images.Then,this paper uses linear regression analysis to verify the strong linear correlation between the adjacent bands of hyperspectral remote sensing images,and then takes advantage of the geodesic distance to verify its significant nonlinear properties,thus demonstrating that the hyperspectral remote sensing images are nonlinear manifolds in high dimensional space.Finally,this paper uses k-means clustering method to verify the certain generalized clustering characteristics of remote sensing images.To sum up,these intrinsic characteristics will provide the theoretical support for the construction of data intrinsic structure driven deep learning methods in this paper.Secondly,in order to overcome the spectral variability and the great overlapping of the spectral bands in the hyperspectral images,this paper proposes a statistical property driven deep learning method for remote sensing image classification to fully utilize the reflectivity structure information of the image by modeling the spectral information with the statistical characteristics.Specifically,this paper uses probabilistic models to model different classes of hyperspectral data,and then constructs the optimal equations by Fisher discriminant criteria and diversification criteria.Finally,the multivariate statistical analysis theory is used to statistically analyze the optimal equations for the final statistical loss.The probabilistic model and the statistical analysis have been applied to model the statistical property of the hyperspectral image,so that the intrinsic statistical characteristics of the data can be analyzed by the limited number of training samples.The statistical loss proposed in this paper especially shows stable performance for the classification problems under very small number of training samples.Thirdly,to overcome the low training efficiency of measuring the distances between the samples in hyperspectral images under the European metric,this paper proposes a intrinsic manifold structure driven deep learning method to improve the performance of the deep learning model by constructing the geometric structure of the data.Specifically,this paper extracts the intrinsic manifold structure of the hyperspectral image using the geodesic distance and the clustering distribution law,and then constructs the optimal equations by maintaining the intrinsic manifold structure in the low dimensional feature space,and finally through the diversity-promoting methods to increase the variance between the sub-classes of different classes,deep manifold embedding based deep learning method is constructed.Experiments on three hyperspectral data validate the effectiveness of the deep manifold embedding method.Furthermore,the paper deeply analyzes the advantages and disadvantages between the deep manifold embedding method and the statistical loss.The results have shown that the performance of the deep manifold embedding method outperforms the statistical loss when more than 120 training samples per class are used for experiments over Pavia University data.Besides,the deep manifold embedding method is better than the statistical loss when using more than 80 training samples per class for experiments over Salinas Scene data.Fourthly,in order to solve the problem that few labelled samples cannot build the class model effectively,this paper proposes a supervised learning method based on generalized clustering property for remote sensing image classification which can utilize the class information and ensure the good classification performance.Specifically,this paper develops a center point-based structured metric learning method by embedding the generalized clustering property into the structured loss and using the data intrinsic clustering property to enhance the representational ability of the deep model for remote sensing images.Experiments over three datasets,namely the UCM,Brazilian,and Google dataset,demonstrate that the use of the generalized clustering property within the remote sensing images can significantly decrease the classification errors of the deep learning models.Finally,to overcome the problem of the limited number of labelled training samples in remote sensing images,this paper proposes a generalized clustering property driven unsupervised transfer learning for remote sensing image classification,which can learn the deep model without the labelled samples.Specifically,this paper uses the Image Net dataset to pre-train the model,so that the model can obtain certain useful information from the source domain through transfer learning.Then,by constructing the diversified pseudo-center loss and the softmax loss based on the pseudo-classes,the generalized clustering property driven unsupervised transfer learning method is developed.Experiments on three remote sensing image datasets validate that the unsupervised transfer learning method driven by the generalized clustering property can achieve or almost reach the performance of supervised learning,so that the deep learning process can partially get rid of the labelled sample dependence.