Research On Multi-source Remote Sensing Data Fusion Based On Deep Learning

Remote sensing images have important applications in agriculture,environmental monitoring,military and other fields,but single-source remote sensing images have some defects in spatial resolution,spectral resolution and other aspects.Based on this,multi-source remote sensing data fusion research is proposed to comprehensively improve the effect of remote sensing image application.In this dissertation,the fusion of hyperspectral image and visible image is selected as the main research direction,which is divided into two aspects:first,the use of visible image assisted hyperspectral image for super resolution reconstruction.The second is to use visible image to assist hyperspectral image for target detection.The super resolution reconstruction technique of hyperspectral image is a method to improve the spatial resolution of hyperspectral image.However,most of the existing methods are based on the strict registration of hyperspectral and visible images,and the spectral information recovery is not good after reconstruction,which will introduce spectral distortion.Based on this,this dissertation proposes a hyperspectral reconstruction algorithm based on double-stream preprocessing.Firstly,it proposes the preprocessing of multi-scale feature extraction for visible light to solve the problem of requiring advance registration,and reconstruction of hyperspectral by cross-sampling up and down respectively to solve the problem of poor spectral information recovery.At the same time,the squeeze and excitation module are applied to the hyperspectral superresolution reconstruction task for the first time to build the fusion module.By combining the information between channels,the channel attention is enhanced to realize the spectral dimension calibration and improve the network performance.Compared with other methods,the method proposed in this dissertation has advantages in four evaluation indexes.At present,most of the advanced detection methods of hyperspectral targets pay attention to spectral information but ignore spatial information,which leads to some problems such as different spectrum of the same object,foreign body of the same spectrum,mixed pixel and so on.In this dissertation,a target detection algorithm based on the fusion of hyperspectral data and visible image data is proposed.The visible image is used as input to assist hyperspectral image for target detection,so as to obtain better edge information.In addition,the whole image is used as input,thus utilizing both spectral information and spatial information.A multi-level feature pyramid network structure is designed to prevent the disappearance of features,and supervision is carried out at each level to obtain better detection results.A multi-size convolution kernel object detection network structure is built to detect objects better.The precision of target detection is improved by adding edge supervision.The method proposed in this dissertation is tested on two data sets with four evaluation indexes,and obtained good results.