Research On Scene Classification Of Remote Sensing Images Based On Deep Learning

Remote sensing image scene classification has always been an important research content in the field of remote sensing,widely used in many fields,and has important significance for the development of remote sensing technology.Remote sensing image scene classification is to distinguish remote sensing images based on the differences in feature information of different categories of remote sensing images.Compared with general images,the texture and color feature richness of remote sensing images is higher,and there are also problems such as complex background and many types.Therefore,there are certain challenges in the classification of remote sensing images.This paper proposes a remote sensing image scene classification method based on the combination of multi-level features,combining remote sensing images with bottom features,middle features,and high features to jointly express the semantics of remote sensing scenes,and training a classifier to construct a remote sensing image scene classification model.Different levels of features complement each other with feature information to improve classification accuracy.Firstly,extract the bottom GIST features of the remote sensing image and combine the GIST features of multiple color spaces to enhance the color information of the remote sensing image.Secondly,extract the middle bag-of-words features of the remote sensing image,calculate the gray-scale image,the maximum image,the minimum image and the average image of the remote sensing image,extract the bag-of-words features from the four images,and combining optimal bag-of-words features of image under different clustering numbers to improve feature expression ability.Then extract the high features of the remote sensing image,use the convolutional neural network to train the remote sensing data sets and extract the deep features,introduce the transfer learning strategy into the network training task,save training time and improve the model performance.Finally,the image features of multiple levels are normalized to the same range to reduce the difference between different levels of feature data,and the combined features are input to the classifier for classification.The UCM Land-Use and WHU-RS19 two remote sensing data sets are used to verify the proposed method.The experimental results show that the proposed method has improved classification accuracy compared with other existing methods,and has better advantages.