Multi-modal Remote Sensing Image Fusion Classification Method With Small Sample Size

With the development of remote sensing technology,the spatial and temporal resolution of remote sensing images has been continuously improved,which has led to a large number of remote sensing image data analysis techniques.As an important branch of remote sensing image research,remote sensing image classification technology has been widely used in marine monitoring,resource exploration,agricultural monitoring,military identification and other fields to provide technical support for resource planning,management and decision-making.The traditional neural network remote sensing image classification method can not meet the requirements of local association and parameter sharing.The image classification method based on convolutional neural network is a neural network with local perceptual structure,which can take image data as input and extract local-to-global features layer by layer,which can effectively classify images.The training of the traditional convolutional neural network model requires a large number of training data sets,but it is difficult to obtain a large number of labeled remote sensing image samples in practice,so that the remote sensing images exhibit the characteristics of small samples.Up to now,there is still a lack of a convolutional neural network that can be effectively applied to small sample remote sensing image classification.Based on this,combined with the multi-modality of remote sensing image,this paper studies a multi-modal multi-modal fusion method to achieve high-precision classification of remote sensing images.In order to achieve high-precision classification of multi-modal remote sensing images with small samples,the paper has carried out in-depth research on multi-modal fusion strategies.In view of the difficult training of deep convolutional networks,combined with migration learning methods,many migration-based learning methods are proposed.Modal fusion classification method.The main research contents include:(1)A small sample multi-modal remote sensing image serial fusion classification methodThe traditional multi-modal remote sensing image fusion classification method fails to fully consider the correlation between modes and the interference between modes.Classification).Among them,the high-level features of different modes are important steps to improve the classification accuracy.In this paper,the convolutional neural network is used to extract the high-level features of two different spatial resolution remote sensing images,and the fusion is constructed by a serial fusion strategy with two highlevel feature correlations.Finally,the high-level feature training based on the strategy fusion is obtained.The entire classification model.Experiments show that the accuracy of the MRSIC model for remote sensing image classification is 90.5%.(2)A small sample multi-modal remote sensing image parallel fusion classification methodThe serial fusion classification model requires that the test data of the two modes must be input into the corresponding classification model,which has high requirements on the testers' professional skills and seriously affects the application and promotion of MRSIC.To this end,a parallel fusion fusion classification method is proposed,which is denoted as N-MRSIC(New Multi-modal Remote Sensing Image Classification).This method achieves high applicability by the method of classifying the test data of a single modality.Experiments show that N-MRSIC greatly reduces the pre-processing of test data,reduces the technical requirements of the classification model for users,and improves the high applicability and classification accuracy of the classification method.(3)Proposed multi-modal remote sensing image fusion classification method based on migration learningFor a wide variety of classification applications,the learning ability and image decomposition accuracy of the shallow CNN(Convolution Neural Network)is difficult to guarantee.To this end,a complex CNN classification model is proposed,but complex CNN training is difficult.Based on this,a T-MRSIC(Tranfer-learning Multi-modal Remote Sensing Image Classification)improved with the migration technique is proposed.Firstly,in the complex CNN classification model constructed,the model trained on the large data set(ImageNet)is introduced,the parameters of the first few layers are fixed,the last one to two layers are modified,and the method of retraining the classification model with its own data is used.Experiments show that under the challenge of less training samples and more types of samples,the accuracy of using only the migrated single-modal classification model is 90.3%,and the classification accuracy of T-MRSIC can reach 93.3%.Through the research of the above contents,the thesis has obtained certain research results.(1)A serial fusion classification model with anti-interference is proposed,which makes the multi-modal remote sensing image information of small samples complementary;(2)To reduce the classification model The test data must exist at the same time and the requirements of the tester's professional knowledge,and the application should promote the highly applicable parallel fusion classification model.(3)Facing the complex CNN model training problem,introduce the parametric model migration technology to realize the complex CNN of small samples.Highly accurate classification of models.