Research On Remote Sensing Image Scene Classification Based On Deep Learning

With the rapid development of satellite and aircraft observation technology,data acquisition of high-resolution satellite remote sensing images or airborne SAR remote sensing images has become easier.This has led to an important need for intelligent observation of the Earth through remote sensing image scene classification.The thesis studies the classification of remote sensing image scenes based on the deep learning method.The main work is as follows:The traditional high-resolution satellite remote sensing image scene classification method mainly relies on the artificially extracted middle and low-level features and can not make good use of image-rich scene information.To solve this problem,a band-based feature fusion and GL-CNN are proposed.Learning a convolutional neural network)classification method.Firstly,the high and low frequency sub-bands of the image are extracted by non-downsampling wavelet transform,then the high frequency sub-band is fused to obtain the fused high-frequency sub-band,and then the harmonized high-frequency sub-band is analyzed by the stationary interval analysis of the angular distribution of the energy distribution curve.The sample is fused with the samples of the low-frequency sub-band,and finally the convolutional neural network is automatically extracted to extract the high-level features contained in the high-low frequency sub-band of the image to implement scene classification.Experiments on UCM_LandUse 21 data show that the classification accuracy rate is 94.52%,which is significantly higher than previous algorithms.In order to further improve the feature learning ability of convolutional neural networks in scene classification,a remote sensing image scene classification method based on multi-scale spatial statistical modeling and feature recalibration network is proposed.Firstly,multi-scale Omnidirectional Gauss Derivative Filter(MOGDF),Gray Gradient Co-occurrence Matrix(GLGCM)and Gabor transform are used to obtain multi-scale spatial statistical features of remote sensing images.Then,high-level feature space relationships are introduced.A feature recalibration module with depth separable convolution,which acquires correlation between feature channels by suppressing and exciting mechanisms and performs weight screening on multi-feature inputs,and finally completes nesting by feature recalibration module and convolutional neural network Remote sensing image scene classification.The scene classification test of UCM_LandUse and 11 SAR data is carried out,and the classification accuracy rate is 96.67%and 98.18%respectively.The experiment shows that the method can significantly enhance the generalization ability of the network and improve the classification accuracy.Because the traditional feature extraction method mainly relies on manual design.It not only requires a lot of theoretical knowledge and a solid professional foundation,but these features are based on the underlying visual features of the target and cannot fully characterize the essential attributes of the target.In recent years,the emergence of deep learning provides another solution for remote sensing image scene classification.As a machine learning method based on sample data representation learning,deep learning can effectively solve many problems in traditional image target recognition methods.Ability to automatically extract higher-level abstract features of images,reducing manual intervention.