Research On Pan-sharpening Method Based On Multi-scale Feature Depth Neural Network

With the rapid development of satellite remote sensing system,remote sensing images are widely used in disaster early warning,forest vegetation monitoring,military detection,etc.However,due to the limitation of technology,a single sensor cannot obtain remote sensing images with high resolution spectral and spatial domains at the same time.Therefore,in practical applications,the spectral and spatial components need to be processed in combination to obtain multispectral images with high spatial resolution through panchromatic sharpening.In recent years deep learning has been widely used for panchromatic sharpening tasks,with remarkable results.However,most methods fail to effectively seek mapping relationships between images,or lack well-designed fusion rules for the overall network structure.The panchromatic sharpening method is studied in this thesis,and obtains high-quality remote sensing fusion images based on attention mechanism,multi-scale feature fusion and parallel interaction network.The main research contents are as follows.1.A panchromatic sharpening method based on a multiscale delayed channel attention network is proposed.The method designs a positive feedback module,which maps the error at each stage through an error feedback mechanism,thus positively feeding back the feature extraction quality of the network.In addition,a multi-scale feature fusion module is designed to improve the information extraction capability of the network by fusing feature information under different visual fields.Meanwhile,a delayed channel attention mechanism is proposed to obtain the correlation between low-frequency information and high-frequency information through adaptive learning,giving different weights to high-frequency information to make the network more flexible in processing different types of information.From the obtained experimental results,the method obtained high quality fused images.2.A method of panchromatic sharpening with parallel interaction network as the main structure is proposed.The method combines the domain knowledge of panchromatic sharpening with the parallel input of subnetworks to reduce the network depth required for the establishment of multi-resolution subnetworks.The method obtains feature information under different fields of view through multi-resolution subnets and improves the representation of high-resolution features by multi-scale low-resolution features of the same depth and similar levels.A delayed channel attention module is designed based on the delayed channel attention mechanism.The overfitting representation is avoided while improving the feature selection ability of the network.Experimental results show that the proposed method is highly competitive in both quality assessment and visual aspects.