Multi-Temporal Remote Sensing Image Registration Using Deep Convolutional Features

Registration of multi-temporal remote sensing images has been widely applied in military and civilian fields, such as ground target identification, urban development assessment, and geographic change assessment. Ground surface change challenges feature point detection in amount and quality, which is a common dilemma faced by feature-based registration algorithms. Under severe appearance variation, detected feature points may contain a large proportion of outliers, whereas inliers may be inadequate and unevenly distributed. This paper presents a convolutional neural network (CNN) feature-based multi-temporal remote sensing image registration method with two key contributions: (i) we use a CNN to generate robust multi-scale feature descriptors and (ii) we design a gradually increasing selection of inliers to improve the robustness of feature point registration. Extensive experiments on feature matching and image registration are performed over a multi-temporal satellite image data set and a multi-temporal unmanned aerial vehicle image dataset. Our method outperforms four state-of-the-art methods in most scenarios.