Research On Remote Sensing Image Classification Method Based On Deep Transfer Learning

Remote sensing image classification is a work of great significance in engineering practice and plays an irreplaceable role in many fields such as civil use,scientific research and military affairs in modern society.With the continuous launch of remote sensing satellites,the amount of remote sensing image data is still increasing,and the classification method of remote sensing image is also developing rapidly.This project takes the deep learning classification of multi-spectral Landsat8 remote sensing images with small samples as the background,and studies the classification methods and optimization strategies in combination with the idea of transfer learning.The research contents of the project mainly cover the following points:First of all,according to data pretreatment for deep learning problem,the first step to obtain the correct image spectral and spatial information,with original multispectral images in the study area as the research object,through visual interpretation to determine the object type and spatial distribution,then studied the radiation calibration and atmospheric correction method,in order to obtain correct spectral information.The method of image fusion is studied to improve the spatial resolution,the index of evaluating image fusion is given,and the validity of image fusion is verified by experiments.The second step is to enhance the image through band synthesis,so that all kinds of features are bright in color and easy to distinguish.The evaluation index of band synthesis is given,and an optimal synthesis method suitable for the research area is obtained through experiments.The third step is to make a deep learning-oriented data set.The method to enhance the small sample data and the index to evaluate the validity of the sample data set are given,and the validity of the experimental data set is verified.Secondly,in view of the deep learning method in the small sample training data set on difficult problems,introduced the migration method of study,the migration on large data sets to estimate of the trained network target data set,on the fine-tuning,and through the performance comparison before and after the experiment analyzed the fine-tuning,verify the validity of the fine-tuning of network performance improvement.Secondly,aiming at the problem of cross-domain difference in the process of migration,a normalized adaptive structure is introduced to reduce it.On this basis,a new migration network structure is designed,and the effectiveness of the structure in reducing cross-domain difference is verified by experiments.To solve the problem that single transfer feature richness is not enough,a multi-feature fusion strategy is designed by introducing a spatial pyramid structure to extract features at multiple levels.This method first calculates the classification accuracy of the feature to be fused on the test set,and then determines the weight of feature fusion by the ratio of these precision values.Finally,a comparative experiment was conducted.Firstly,the artificial feature method is used to classify the target set,and the classification evaluation index is compared with the method in this paper,so as to verify that the extraction method is more representative than the artificial feature.Classification was performed on other small sample data sets to verify the generality of the experimental migration network and to compare it with other migration networks.Experiments on other small sample sets verify that the designed feature fusion strategy also improves the classification accuracy and migration features of other data sets,and the effectiveness of this method is verified by comparing with the feature fusion strategy in the reference experiment.