Research On Registration For Multisource High-Resolution Remote Sensing Images

As highly destructive natural disasters,large earthquakes have the characteristics of extensive damages,strong suddenness and serious losses.Carring out disaster damage assessment and disaster relief and rescue quickly after the earthquake is one of the effective ways to reduce the losses.In remote sensing earthquake disaster assessment,multi-source remote sensing image registration is an important step.With the development of RS technology,different satellite sensors can provide multispectral,multitemporal,multiplatform,and multiresolution RS images.The coregistration of multisource high-resolution RS images has been a research focus in RS image processing.Local deformations may occur on high-resolution RS images acquired after severe earthquakes,which increases the difficulty of multi-source remote sensing registration.For images with complex terrain and surface deformations caused by disaster-damages,traditional methods cannot meet the requirements of the accuracy and efficiency of the image registration.To solve this problem,from the perspective of features,two automatic registration methods for multisource high-resolution satellite remote sensing images from the perspective of features were proposed in this paper.Two pairs of GaoFen satellite images and three pairs of pre-and post-earthquake high-resolution satellite images were used to verify the advantages of the proposed methods in terms of the number and distribution of tie points,running time and registration accuracy.The main achievements of this research are as follows:1.The traditional remote sensing image registration methods are time-consuming and low precision,which can hardly meet the requirements of accuracy and efficiency simultaneously.A fast automatic registration method based on Shi_Tomasi and scale-invariant feature transform(SIFT)algorithm was proposed for multisource high-resolution remote sensing images(SSR).In this method,the combination of the Shi_Tomasi corner detection algorithm and SIFT to detect tie points from image patches obtained by an image partition strategy considering geographic information constraints.Then,random sample consistency(RANSAC)and control point homogenization were used to remove the error and redundant matching points,respectively.The experimental results show that compared with the classical SIFT method(SIFT)and the SIFT method using the same image segmentation strategy(Patch-SIFT),this method has the advantages of high registration accuracy,short running time,many matching points and even distribution.For domestic high-resolution remote sensing images,the running time is improved by 3.24 times and 2.01 times respectively,and the registration accuracy is improved by 0.26 pixels and 0.25 pixels.For post-earthquake high-resolution satellite remote sensing image registration,the advantages are more significant,with the operation time increased by 11.17 times and 2.48 times respectively,and the registration accuracy increased by 4.12 pixels and 0.87 pixels,which can better meet the requirements of post-earthquake disaster assessment for remote sensing image registration accuracy and efficiency.2.In order to improve the accuracy of matching features,an automatic registration method based on convolutional neural network(CNN)and SIFT was also proposed for multisource high-resolution remote sensing images.Firstly,a registration sample set was constructed from high-resolution satellite remote sensing images.Then,basing on the initial model parameters obtained using the ImageNet database,the training model was obtained by transferred learning and fine-tuned by combining with the sample set.Finally,the fused feature obtained through the combination of the CNN features and SIFT features was used to detect tie points.Specifically,the VGG16 and ResNet50 networks were employed to construct three fused features,which including SIFT+Fc6,SIFT+Fc7,and SIFT+ResNet50.The experimental results showed that the new method using fused features can achieve higher accuracies and more tie points.Compared with SSR method,Patch-SIFT and SIFT,the registration accuracy of remote sensing image is improved,which is 0.11,0.67 and 2.30 pixels respectively.However,how to improve the efficiency of the new method is one of the works to be further studied.