This is a Preprint and has not been peer reviewed. This is version 2 of this Preprint.
Evaluating Modern Tools and Large Language Models for Digitising Sea-Level Data from Marigrams
Downloads
Authors
Abstract
Numerous graphical data digitisation tools exist; however, few have been applied to the recovery of historical sea-level marigrams (tidal charts). Using an early twentieth-century marigram from the Clyde Estuary, Scotland, containing one week of sea level observations, we evaluated the performance of 11 digitisation tools and 2 large language models (LLMs) in digitising marigrams. A multi-user testing framework was used to assess the accuracy and speed of each approach. Both LLMs tested, ChatGPT5.5 and Gemini 3.1Pro, were unsuitable for this application under the tested conditions, because they show errors exceeding 10% of the observed tidal range. Conventional digitisation tools performed substantially better, most digitisation tools successfully recovered sea-level data with errors below 1%. Despite this accuracy, these tools remained dependent on substantial human input. Even the fastest approach, DigitGraph, required on average 29 minutes to recover one week of sea-level data. Scaled to a century-long record, this would require more than one working year to complete, making application to large datasets challenging. Consequently, further development of dedicated, automated digitisation algorithms remains necessary.
DOI
https://doi.org/10.31223/X5P51Q
Subjects
Life Sciences
Keywords
Marigram, Digitisation, Large language model, Data recovery
Dates
Published: 2026-09-28 09:12
Last Updated: 2026-09-30 00:07
Older Versions
License
CC BY Attribution 4.0 International
Additional Metadata
Conflict of interest statement:
None
Data Availability:
The data that support the findings of this study are openly available in Evaluating Modern Tools and Large Language Models for Digitising Sea-Level Data from Marigrams at https://zenodo.org/records/21921305.
Metrics
Views: 269
Downloads: 75
Comment #347 Frédéric Pons @ 2026-10-02 02:40
Thank you for this paper and these tests.
I would like to offer two comments on the use of NUNIEAU, followed by a discussion. I am attaching an old document about NUNIEAU (https://cerema.box.com/s/8ddjl3h6tumkww8typ0uq8hdk1w2b6w6) to illustrate some examples.
Lines 286–288:
"NUNIEAU also produced comparatively high RMSE values. Its workflow required users to recalibrate the graph after each trace, increasing both user interaction and completion time while introducing calibration inconsistencies between digitised days."
Response: There is actually no need to recalibrate the graph before each trace: for a given image, several calibration files can be associated with it. Once the first graph has been processed, moving on to the second one simply requires incrementing the digitisation number associated with the sheet and, for a daily marigram, applying a 24-hour offset.
There is also a dedicated module for injecting the predicted tide and cleaning the data around this predicted tide, which should in principle save a great deal of time. The first month can be difficult, as one needs to build a sample to obtain a predicted tide if none is available for the site, but after that it becomes a great help for understanding the continuity between graphs (attached document, page 16).
In NUNIEAU, there is an initial set-up time to start a job; after that, if the colour is the same, if the sheets are the same, and if the same scanner is used, many steps can proceed much faster.
Lines 294–295:
"Although the trace tools were generally less accurate than manual digitisation and the point-and-click tools,"
Response: I cannot speak for the other tools, but NUNIEAU allows the user to go back to a previously defined calibration, which is a guarantee of quality — unlike manual work, which cannot be verified by peers. It is also possible to carry out tidal analyses and to modify the calibration data a posteriori in order to correct issues of scale or time referencing. This is very useful in low-tidal-range areas.
Discussion
1. On these topics — and this is the spirit of NUNIEAU (and perhaps of other tools) — it seems essential to me that a calibration can always be revisited. Digitisation is a tedious task, so it must be possible for the work to be checked by someone external. This also makes it possible to outsource this type of work to external providers and then validate their output.
2. You describe digitising a line or a point. NUNIEAU works on the min/max/mean of the signal, which makes it possible, beyond the mean sea-level requirement, to analyse possible oscillations of the sea level at short time steps — these are not swell, but not-so-simple phenomena such as wave groups (attached document, page 19).
3. I do not think that using an LLM directly is efficient. Instead, one should take the pattern-detection capabilities of LLMs and code them into digitisation tools, after a pre-identification of a narrower area — in your case, working only on a zone around the predicted tide of each recorded day.
We also have a small community (tide, rain and river) on this topic at the following link: https://www.expertises-territoires.fr/jcms/pl1_69072/fr/nunieau-logiciel-de-numerisation-des-enregistrements-graphiques-de-niveaux-d-eau