Research Of Remote Sensing Image Scene Classification Based On Multi-objective Siamese Convolutional Neural Networks

Thanks to the rapid development of remote sensing techniques in recent years,the scale of remote sensing image is gradually expanding,and the image itself is providing more and more rich information.However,how to deal with the rich information and effectively use it in such major applications as natural disaster detection,environmental detection and urban planning is a serious challenge for the remote sensing community.Remote sensing image scene classification,as a key component of remote sensing field,has always been a research focus in the academic field.The main process of scene classification is to distinguish images with similar scene features from multiple images according to the extracted image features,and assign correct category labels to these images.At present,a large number of remote sensing scene classification methods have been proposed one after another,mainly based on deep learning method.Although scholars at home and abroad have done a lot of research on deep learning models in the field of remote sensing scene classification,the current research still has the following deficiencies:(1)few types of remote sensing image data sets and the small number of image annotation samples,which limits the full potential of deep learning methods;(2)due to the influence of sensor type,wavelength,the diverse shooting angles of images and different illumination conditions and other factors,the spatial resolution and spectral resolution of remote sensing image are uneven and remote sensing images have different appearances,which increases the difficulty of scene classification;(3)the remote sensing image has the characteristics of within-class diversity and between-class similarity,which increases the difficulty of extracting robust features for the existing methods.In order to solve the above problems,a Siamese convolutional neural network is introduced in this paper to carry out a series of research on remote sensing image scene classification.The main research contents of this paper include:1、This paper proposes a Siamese convolutional neural network model for remote sensing scene classification.The whole model inputs an image pair from the same scene category or different categories.The model not only classifies the scene of the input image,but also evaluates the similarity of the input image pair.In addition,the model also introduces a metric learning mechanism to learn the feature representation extracted from convolutional neural networks,so that the images of the same category in the feature space are as close as possible and the images of different categories are as far away from each other as possible,further improving the accuracy of the scene classification of remote sensing images by Siamese convolution neural networks.2、This paper proposes a method of rotation-invariant feature learning and joint decision-making for remote sensing scene classification.The recognition model and verification model are classic models of convolutional neural networks.In feature recognition,the recognition model will give the class probability of image,while in feature comparison,the verification model will give the class similarity probability of image pairs.These two probabilities are introduced into Bayesian rule for joint decision classification of remote sensing images.For the problem of remote sensing image scene classification,the existing scale of remote sensing datasets is not enough to train an accurate and robust classifier.In order to solve this problem,the data augmentation method of random rotation will be used for the remote sensing data set to expand its scale,so that the Siamese convolutional neural network model can learn more robust feature expression,improving the accuracy of scene classification.3、This paper proposes a multi-objective residual network pruning model based on multi-objective evolutionary algorithms.The over parameterization of convolutional neural networks leads to the problems of high calculation cost and large storage capacity.It is found that the parameters of convolutional filters in the convolutional neural network are usually less than those in the fully connected layer,but they occupy most of the computation of the network.It can be seen that pruning the filter can reduce the complexity of the network to a certain extent.In this paper,a random pruning strategy is used to prune the Siamese residual network,and the pruning problem is modeled as a multi-objective optimization problem.The optimization goal is to improve the accuracy of scene classification and reduce the complexity of the network.Therefore,multi-objective evolutionary algorithm is introduced in the model to achieve the best balance between classification accuracy and network complexity.