Multi-Label Classification Of Remote Sensing Image Based On Multi-Scale Convolutional Neural Network

With the rapid development of earth observation systems such as satellites,a large number of remote sensing images have been produced,but many of them cannot be effectively used.Therefore,managing them has become an urgent problem,and classification is an effective way.Remote sensing images have rich semantic information and complex surface features.Generally,an image contains multiple types of surface features.Therefore,multi-label classification is helpful for understanding and managing the image,and it can be found from the image data as needed.An image of a certain type of label.Based on the convolutional neural network and the introduction of multi-scale,this paper proposes a multi-scale classification algorithm for remote sensing image based on multi-scale convolutional neural network.This paper studies the multi-label classification of remote sensing images based on multi-scale features and multi-scale images.The AlexNet network is first modified to meet the needs of multi-label classification tasks,and then optimized by two multi-scale features and one multi-scale image.The multi-label classification of remote sensing images based on multi-scale feature fusion is studied.The multi-scale feature fusion network is represented by U-Net,so it is modified first,and based on this,a recurrent neural network is introduced to extract the correlation of the tags;the impact of this paper is studied at the end There are two factors affecting the accuracy of multi-label classification of remote sensing images.The results show that there are two influencing factors,namely the number of labels and the complexity of label features.The main research conclusions of this paper are as follows:(1)The number of tags and the complexity of tag features both have a certain effect on the score.In the classification results of the two methods in this paper,the average score of the six types of tags with a larger number of tags is higher than that of other tags by 0.022,0.011,respectively.It shows that there is a certain relationship between the score and the number of tags;the number of tags in the sea tag and the tank tag are basically the same,but in the comparison of the scores in the two methods in this article,the former is 0.307 and 0.327 higher.It shows that the score is also highly related to the complexity of the label features.(2)The use of multi-scale features and multi-scale images can effectively improve the accuracy of multi-label classification.In the experiments,multi-scale features increase the overall score by 0.032,and multi-scale images increase the overall score by 0.048,proving that they have acquired more valuable features in multi-label classification tasks,which can effectively improve the accuracy of multi-label classification.(3)Multi-scale feature fusion is operable for multi-label classification tasks.Compared with the method based on multi-scale features and multi-scale images,the overall score of U-Net with multi-scale feature fusion is 0.006 higher,indicating that U-Net can be used not only for image segmentation,but also for multi-label classification of remote sensing images.It is proved that the fusion of features at different levels is conducive to the performance improvement of multi-label classification tasks and is operable.(4)Bidirectional recurrent neural network can effectively mine the label correlation in multi-label classification.After the introduction of the two-way recurrent neural network,the overall score increased by 0.005,and the scores of the ship and water body tags increased by 0.038 and 0.054,respectively.It proves that the two-way recurrent neural network can consider the interdependence between the tags and can effectively mine the tag correlation in multi-label classification.