MSGFlow-Net: Multi-Source Guided Flow Network for Heterogeneous Remote Sensing Image Registration test datase

The Landsat8-Sentinel2 dataset consists of 464 image pairs (256×256 pixels) derived from Landsat-8 and Sentinel-2 surface reflectance imagery covering 5,129.6 km² in Beijing, composited over June–December 2024 using cloud masking and median statistics on Google Earth Engine. The GF1-2 dataset consists of 1,121 image pairs (256×256 pixels) built from GF-1 and GF-2 high-resolution imagery covering 552.25 km², acquired on 1 August 2024 and preprocessed through radiometric calibration, atmospheric correction, and orthorectification. For both datasets, manually defined transformations were applied to generate perturbed source images and displacement fields, enabling controlled evaluation under optical–optical settings. To further assess real-world performance, the Landsat8-Sentinel2-R and GF1-2-R datasets preserve the natural geometric discrepancies inherent in cross-sensor acquisition, each comprising two original images and one manually registered reference as ground truth. This design enables systematic evaluation of registration performance under both standardized and practical cross-sensor, cross-resolution conditions.