Multi-stage matching for different sources of geospatial images based on local image self-similarity

The combined use of multi-source geospatial imagery is crucial for many applications, including image registration and three-dimensional reconstruction. However, complex geometric transformation and nonlinear radiation distortion are always present in different sources of geospatial images, significantly increasing the difficulty of effective matching. To address this, we introduce a robust method for matching multimodal geospatial data by integrating salient image features and dense templates to enhance matching accuracy. Our method begins by extracting local structural information from each pixel using the local self-similarity (LSS) model. We then construct a rotation-invariant feature descriptor that is combined with scale space for coarse matching with geometric invariance. After that, we resample the sense image and conduct feature matching on the reference image and the sampled sense image. Finally, a dense image template is designed and the matching accuracy is refined using the results obtained from feature matching. Experimental results conducted on various types of multimodal geospatial images, including optical, infrared, synthetic-aperture radar, and digital maps, demonstrate that our method surpasses multiple state-of-the-art matching techniques by achieving higher correct match counts, improved precision and recall, and lower root mean square error. These findings validate the effectiveness of our approach in achieving robust and accurate matching across diverse multi-source geospatial datasets.