Historical aerial imagery provides valuable data from regions and periods with limited geospatial information. A common method to utilize this data is through the generation of ortho-photos and 3D models using Structure-from-Motion (SfM) techniques. However, many of these images were scanned decades after their acquisition and require geometric calibration, along with internal and external camera parameter estimation, for accurate reconstruction. Manual identification of key features, such as fiducial marks and text annotations, is labour-intensive, while existing automated methods struggle with poor-quality datasets. This paper presents an automated workflow that combines computer vision and machine learning techniques to detect and extract these key features from historical aerial images. To address challenges related to image quality, we also introduce estimation protocols that compensate for missing or unreliable detections by leveraging redundancy across multiple flight paths. The methodology was evaluated on the TMA (Trimetrogon Aerial) archive, a collection of historical images from the Antarctic Peninsula. Our test dataset comprised over 7000 images from 20 different flight paths. The workflow demonstrated high success rates in detecting and extracting fiducial marks, image subsets, and textual annotations. Approximately 70% of the images provided usable focal length data, while fiducial mark detection exhibited high accuracy except in cases of severe scanning artifacts. Altitude data extraction proved to be the most challenging, with successful results in only 15% of images due to degraded altimeter readings. Despite these limitations, the automated workflow effectively estimated missing parameters, ensuring robust image reconstruction across flight paths. The code for this workflow is open-source and publicly available on GitHub at https://github.com/fdahle/hist_meta_extraction.

Satellite-derived bathymetry (SDB) provides a cost-effective solution for coastal mapping, but challenges remain in model interpretability and uncertainty quantification. This study investigates the applicability of the least-squares-based deep learning (LSBDL) framework for SDB, leveraging its hybrid structure that integrates neural networks with the available least-squares theory to enhance model transparency. ICESat-2 photon-counting LiDAR was used to train depth estimation from Sentinel-2 multispectral imagery over an approximately 30 km × 30 km region of near-coastal bathymetry at Anegada, British Virgin Islands. ICESat-2 provided high-precision depth information, of which 80% were used for training and the remainder for validation. LBSDL depth estimation achieved a root-mean-square error (RMSE) of 2.74 m, representing around 10% of the maximum observed depth, with the best performance in the 2–15 m depth range. These findings demonstrate the potential of LSBDL for interpretable and reliable bathymetric mapping, highlighting ICESat-2 as a globally accessible training and validation source and advancing SDB capabilities for data-sparse coastal regions.