Deep Learning for Satellite Image Matching

Recent advances in deep learning have significantly improved image matching under challenging conditions involving large variations in viewpoint, illumination, and appearance. However, existing datasets, learning frameworks, and evaluation protocols are designed primarily for ground-based imagery acquired using pinhole cameras. As a result, when the trained image matching models are applied to satellite images, you end up with significantly suboptimal performance in image matching. The reasons for the suboptimal performance are several: Satellite images are formed one line at a time as the satellite is in motion, wide variations in viewing angles across orbital tracks, sparse revisit patterns, etc.This thesis addresses these challenges by advancing both the dataset design and the model architecture for satellite image matching. It introduces SatDepth, a novel dataset providing dense ground-truth correspondences tailored for satellite imagery. A scalable processing pipeline is presented, along with a rotation augmentation strategy that enables correspondence discovery under large relative rotations. Benchmarking four state-of-the-art image matching frameworks shows that training on SatDepth improves accuracy and generalization to unseen track angles and areas of interest, achieving up to 40% improvement in precision over models trained on ground-based datasets.Building on this dataset, the thesis proposes EpiMaskXA-Net, a semi-dense image matching network that explicitly incorporates satellite imaging geometry. The model integrates patch-wise affine camera approximations and an epipolar-distance-based masked cross-attention mechanism to restrict correspondence search to geometrically plausible regions, while fine-tuning a foundational pretrained image encoder. Experiments on SatDepth demonstrate state-of-the-art performance, with up to 30% improvement in matching accuracy compared to retrained ground-based models.Overall, this work establishes a unified, geometry-aware framework for satellite image matching, enabling robust image alignment and supporting downstream applications such as sparse 3D reconstruction in remote sensing.