Remote Sensing Image Classification Algorithm Based On Deep Feature Fusion

Remote sensing image classification plays a role that cannot be ignored for today's social development.Therefore,many researchers have been studying in this field to propose more advanced theoretical methods to achieve higher value.Machine learning has become a contemporary research method for remote sensing image classification.At the same time,this method has also been widely used in other major fields.Deep learning,the sub-field of machine learning,whose application field is becoming more and more extensive.Especially in digital image processing,deep learning has achieved great success.Remote sensing image classification is one of the current application scenarios.However,due to the continuous improvement of human requirements for technical methods,research methods must be continuously updated.Therefore,deep learning has always been the main research method for image tasks.Based on Convolutional Neural Network(CNN),this thesis designs a model with good classification effect for several remote sensing image data sets.The following are the main contents of the thesis research:1.Aiming at the problem of partial information loss caused by insufficient feature utilization in remote sensing image scene classification,which further affects the classification accuracy,a remote sensing scene classification algorithm based on dense feature fusion is proposed.Construct Dense Convolutional Network(DenseNet)and extend DenseNet to extract image local features and image global features respectively and use the bag of visual words(BOVW)model coding method for recombination coding to fully express the local deep information of the image.The features extracted from the two parts of the network are linearly weighted and combined with the softmax classifier for classification.Taking use of the local and global feature complementarity to fully extract and use graphic information,which is beneficial to improve the classification accuracy.2.Hyperspectral remote sensing image(HSI)has huge application value with rich data information and high spectral resolution.Convolutional neural networks have had a significanteffect in HSI classification tasks.However,the limited labeled HSI sample makes existing HSI classification methods based CNN usually suffer from small sample size and class imbalance,which is a thorny problem in the current HSI classification task.Therefore,a special CNN architecture is designed in this thesis to find a direction for solving the above problems.A branch structure is introduced on both sides of the traditional network to design a CNN fusion network with a multi-branch structure,that is,a wider and deeper convolutional neural network.This network structure can more effectively extract useful information in the hyperspectral remote sensing images,and the introduction of L2 regularization makes it possible to obtain good results when classifying small sample data sets,thereby improving the generalization ability of the model.