Research On Fusion Algorithm Of High Spatial Resolution Remote Sensing Image For Classification

With the rapid development of earth observation technology,China has launched a series of high spatial resolution remote sensing satellites,such as ZY3,GF1,GF2.High spatial resolution remote sensing satellites can quickly acquire images with rich feature information,which have been widely used in land,resources,environment,disasters,agriculture,forestry,surveying and mapping.However,due to different sensor types,there have different temporal resolution,spatial resolution,and spectral resolution in images acquired by satellites.In practical applications,single-band images have the characteristics that it is difficult to meet the actual needs.It is necessary to fuse images with different spatial resolutions and different spectral bands to obtain more comprehensive images of observation information and provide a better data source for image applications.Image fusion is the pre-processing process of image classification application.Image fusion quality has a great influence on the subsequent image classification.Therefore,this paper is oriented to the application of high spatial resolution image land cover classification,based on domestic ZY3 and GF2 images to study the fusion algorithm.The improved algorithm,the main research contents and results of this paper are as follows:(1)Apply multi-spectral images combined with panchromatic images to obtain multi-spectral images with high spatial resolution,providing a reliable data source for the classification of image land cover.The common IHS transform,PCA transform,Brovey transform,HPF method,SFIM method,Wavelet transform,Contourlet transform and other fusion algorithms are selected to fuse the panchromatic images and multi-spectral images of domestic ZY3 and GF2 satellites,and the fusion images are classified.The suitability of fusion algorithm and ZY3 and GF2 images is evaluated from the aspects of spectral retention,spatial information injection degree and image classification accuracy of fused images.The PCA method has better overall fusion effect on ZY3 images and the spectrum remains better,whose spatial information enhancement and fusion image classification accuracy are superior to other methods.The spatial information of GF2 image is better enhanced than SFIM method,whose spectrum is better and the fusion image has the highest classification accuracy.Its overall fusion effect is better than other methods.(2)Object-oriented classification methods are more suitable than pixel-basedmethods for classification of high spatial resolution images with rich spatial details.The image spatial detail information is an important feature of the object-oriented classification method,which has certain influence on the classification result.In order to provide higher quality fusion images for image classification,this paper applies the algorithm combining super-resolution reconstruction of convolutional neural network and NSCT transform to fuse multi-spectral images and panchromatic images.The algorithm firstly performs super-resolution reconstruction on low-resolution multi-spectral images to maintain spatial spectral characteristics while enhancing spatial information.Then it apply IHS and NSCT transform to fuse panchromatic and the enhanced multi-spectral images.The fusion rule uses multi-spectral image coefficients for low frequencies to maintain spectral characteristics of multispectral images.The high frequency adopts a strategy of maximizing the absolute value to be able to fuse the larger coefficient of the energy,so that the image is more clear.The results show that the overall fusion effect of the proposed algorithm is better than that of the comparison algorithm,which can effectively inject spatial details while retaining spectral features of the fused images more effectively,and obtain better quality fused images.On this basis,based on the object-oriented classification method,the land cover classification of the fusion image is carried out,and five land types such as construction land,bare land,green land,water body and industrial land are extracted.From the aspect of accuracy of terrestrial information extraction,the suitability of fusion algorithm for ground information extraction is discussed,and a fusion algorithm suitable for local information extraction is obtained.