Multispectral Image Fusion And Classification Based On Attention Mechanism And Complex Curvelet Fusion Network

With the development of remote sensing technology,multispectral image is widely used in daily life and military affairs.Multispectral image carries not only the morphological and structural information of ground objects,but also the spectral features of ground objects.Using the multispectral image to classify ground objects is an important research topic in the field of remote sensing image,with significant theoretical value and practical application value.As the amount of information contained in the image increases,traditional image classifications have difficulties in extracting useful classification information and meeting the requirements of practical applications due to their limited ability to mine image features.In the applications of image,deep learning shows powerful ability to express the image features,which becomes a new method for mining deep features and classification information of images.In addition,even though the spectral information is abundant in multispectral images,their spatial resolution is usually low,and the spatial detail information is little,which affects the classification results of multispectral image to a certain extent.A general way to solve this problem is called image fusion,which can gain much more spatial information to improve the classification effect of the images.On the basis of multispectral images,this thesis deeply studies the fusion and classification of multispectral images based on attention mechanism and deep learning.The actual data is used to do the experimental simulation.The main contents are as follows:1.A multispectral image classification based on attention mechanism and DenseNet is proposed.In DenseNet,as the number of feature maps increases gradually,some information redundancy occurs.Among these feature maps,some of them are very useful for classification,while some of them are not so useful for classification.Therefore,a DenseNet based on attention mechanism is proposed,and this network is named as weighted DenseNet.The network can pay more attention to the features that are useful for classification,ignore the features that have little effect on classification,reduce the problem of information redundancy,improve the efficiency of multispectral image classification,and enhance the classification effect.2.A multispectral image classification based on complex weighted DenseNet is proposed.At present,most of neural networks are built on the real number domain.However,studies have shown that complex neural networks are more powerful to express the features and also have stronger robustness.Therefore,the weighted DenseNet is extended to the complex domain,and a complex weighted DenseNet network structure is proposed,which enables the network to extract and express the features of image more effectively,and to ensure the classification effect is more accurate and effective.3.A multispectral image fusion and classification based on complex curvelet fusion network is proposed.By fusing multispectral image and panchromatic image,the advantages of the two images can be combined to obtain more useful information for classification.Since curvelet transform has strong anisotropy,the fusion of the two images by curvelet transform can enhance the ability to express the images.Therefore,a complex curvelet fusion network based on complex weighted densenet combining with curvelet transform is proposed to further improve the classification effect of multispectral images.