Aiming at the problems that remote sensing image features are difficult to extract and the existing image registration framework has low registration accuracy and efficiency, a bidirectional remote sensing image registration method that combines multi-order features and cross-spatial attention is proposed. First, cross-spatial attention is designed to retain multi-scale accurate spatial structure information into channels, and embed it into efficient network blocks to focus on capturing the key information of the image. Secondly, a multi-order feature adaptive fusion module is proposed to be used in feature extraction to adaptively fuse low-order and high-order features to improve the accuracy of registration. Finally, an enhanced feature matching method is designed to analyze the similarity of features more accurately, establish a two-way matching relationship, and use secondary affine transformation to improve the accuracy and reliability of registration. This method achieved 94.0% correct keypoint probability (PCK) on the Aerial Image data set when α=0.05 (α: normalized distance threshold), and the average registration time reached 0.93 seconds. The results show that this method significantly improves the registration accuracy and efficiency of multi-source heterogeneous remote sensing images.