A CNN-Based Framework for Geometric Alignment of Historical and Satellite Imagery

This paper addresses the difficulty of aligning historical aerial photographs with other archive images or modern satellite imagery, which is crucial for applications such as landscape comparison, shoreline erosion monitoring, and urban expansion evaluation. Traditional approaches, such as SIFT for key point identification and RANSAC for feature alignment, suffer with complex geometric transformations. To address these limitations, we present a Machine Learning (ML)-based technique that uses Convolutional Neural Networks (CNNs) for geometric matching. We use a custom-labeled dataset from Google Maps satellite imagery to build paired images with known homograph matrices using controlled transformations such as translations, warps, and rotations. A CNN is trained to predict the homographs between these pairings, allowing for accurate alignment. The suggested CNN-based approach achieved great accuracy for minor translations ( $\le 20$ pixels) with an MSE of 0.0002, but performance decreased as translation magnitude grew. Minor distortions were noticed for moderate translations (40 pixels) with an MSE of 0.0028, while large translations (60 pixels) resulted in severe mistakes (MSE: 0.0735).