Multi-temporal remote sensing image registration based on siamese network

Abstract

Multi-temporal remote sensing image registration aims to find the optimal alignment between images acquired from different times. The complexity and disparity of features in remote sensing images bring great difficulties to image registration. We propose a deep learning method based on the Siamese network to address this problem. Unlike traditional methods doing feature extraction and feature matching separately. We pair patches from sensed and reference images, and directly learn the mapping relationship between those image patch pairs and their matching labels. This end-to-end network architecture helps us optimize the entire network, which is what traditional methods lack. Besides, we use the spatial scale convolution layer in the feature extraction network to improve scale variations' adaptability. Extensive experiments are conducted on a multi-temporal satellite image dataset from google earth. The results of the experiment indicate that our method can obtain more correct matched points and effectively improve the registration accuracy than traditional image registration methods.