Recovering Archival Tide Gauge Data using AI-OCR

Description

Tide gauge records are amongst the oldest and most numerous oceanographic records available. They provide critical information to the study of changes in tidal signals, storminess, storm surge, and mean sea level. Many years, likely thousands of years cumulatively, of paper data remain undigitized across the globe. Traditional manual methods of digitization are slow. Here we test whether an AI-powered optical character recognition (OCR) system can rapidly and accurately digitize images of paper tide records. About 21,000 hourly records were processed in 2-3 hours per model year, approximately 5-10 times faster than typical human-based efforts. The observed error rate of 1 in 129 points was slightly larger than, but comparable to, human error rates of roughly 1 in 160, based on previous digitization efforts. Image quality (primarily focus) was found to be an important determinant of processing accuracy. Our approach greatly expands the potential throughput of oceanographic digitization efforts, likely removes one of the major impediments to digitization and reanalysis of historical records, and thus our understanding of long-term changes caused by natural and anthropogenic processes.

Institutions

• University of South Florida

  Florida, Tampa

Categories

Licence