SOLAR OBSERVATIONS

What are these data?

From 1950 to 2007 every observation was drawn on a paper sheet, with a hand-written table of the sunspot groups (class, number of spots, B, ΔL, L) and the totals g, f and R. These tables were never typed into the database. In October 2026 the group tables of all 15,664 scanned sheets were read by a computer vision model, and each sheet was checked for consistency: number of rows = g, sum of spots = f, R = 10g + f, and the same L₀ on every row.

• Usable sheets (13,361, 85%) passed the count checks. Their daily R agrees with the international sunspot number (SILSO) as closely as the typed observations of 2008–2026 do (correlation 0.95). In 2007, the only year that was also typed, the totals agree on 278 of 283 days.
• Coordinates are reliable only on rows without the “doubtful” mark.
• The other sheets (2,303) are kept with their raw reading, but their numbers are not shown or used: compare with the drawing.
• These are machine readings, not typed records. For any detail that matters, check the drawing, which is linked on every day.

Bulk data for research on request: gozlemevi@istanbul.edu.tr

How were the sheets read?

The AI model

• The tables were read by an open-weight, general-purpose vision-language model: Qwen3.5, 27 billion parameters (Alibaba Qwen team).
• The model was not trained or fine-tuned for this task. The ready-made model was only given written instructions (zero-shot use).
• The model ran 4-bit quantised (Q4_K_M, ~23 GB) in Ollama, on a single NVIDIA RTX 6000 Ada (48 GB) graphics card in the observatory's own server. No image or data was sent to an outside service.
• Total computing time was about 4 days: ~50 hours for the first reading, ~48 hours for the second.

What the model was asked to do

• Two parts were cut from each sheet: the right-hand panel with the group table, and the bottom strip with the totals g, f and R.
• The model was asked to copy the table cell by cell, exactly as written, without interpreting it. Signs, leading zeros, the letters E/W and “?” were kept as they are. Where a value was crossed out and corrected, the new value was taken.
• The column headings were read too, because the table layout changed over the years: polar r, p in the early 1950s, later B, ΔL, L, ΔLF, U.
• The answer was bound to a fixed JSON schema, with temperature 0, so the same image always gives the same reading.
• The model itself did not convert the cells into class, spot count and coordinates; a separate program did that.

Rule-based corrections

• Common misreadings were corrected by rules. Examples: class “?1” read as “21”, “31” or “71”; the letter B read as 8; a hand-written small a read as Q (Ha → HQ); O/D and I/L confused in the last letter of a McIntosh class.
• The relation L = L₀ + ΔL was used to correct L values with a single misread digit. Rows that could not be corrected were marked “doubtful”.
• The raw reading was never deleted. Every correction is stored in the record (field repairs) with its old and new value.

Two-pass reading

• In the first reading 58% of the sheets passed all checks.
• The 6,533 sheets that failed were read a second time. This time the image was enlarged 2×, the model was given a longer context, and the confusions seen in the first reading were added to the instructions.
• In the second reading the model was not shown the earlier reading. So it was not pushed to make the numbers fit the totals, and the checks stayed an independent test of the new reading.
• For each sheet the reading that did better in the checks was kept. In the end 79% of the sheets were fully consistent and 85% usable.

Validation and limits

• 2007 is the only year that was both read and typed. On the fully consistent sheets of that year, compared with the typed record, the class agreed on 95% of the rows, the spot count on 98%, B on 94%, ΔL on 91% and L on 90%.
• The count checks do not check the class letters: the totals can agree even when a class is misread. Allow for an error rate of about 5% in the classes.
• For 1950–2006 there is no typed record to compare with. For these years the accuracy rests only on the consistency checks. Because of different handwriting and the use of Zürich classes, it may differ from that of 2007.

Next step

This catalog will be used as labels to train a neural network that finds sunspot groups in the drawings automatically and gives their McIntosh class. That network has not been trained yet.

Monthly mean R, 1950 → today gold: read from the sheets · navy: typed observations · red dashed: cycle model

Print version of the chart, PDF (ESWW 2026 poster)

Cycle model

The red dashed line is a model of the solar cycle fitted to the monthly means of the chart. Each cycle is described by the profile of Hathaway, Wilson & Reichmann (1994, Solar Physics 151, 177):

F(t) = a (t − t₀)³ / [exp((t − t₀)² / b²) − c]

• t is in months, t₀ is the start of the cycle, a sets the height and b the length of the rise. c = 0.71 is fixed, as in the original paper. The profile rises quickly and decays slowly, as real cycles do.
• One profile was used for each of cycles 18–25 and the profiles were summed, so that the overlapping start and end of neighbouring cycles add up. The 24 numbers were fitted together to all 884 months (1950 – 2026-10). Months with more observing days count more, and single outlying months cannot pull the fit.
• The transcribed sheets and the typed observations were fitted together, in the observatory's own R scale (not converted to the SILSO scale).
• The model follows the data with a correlation of 0.93; the typical scatter of the monthly means around it is 23 in R. The scatter is mostly real: monthly means of 10–25 observing days change a lot from month to month.
• Cycle 18 is in the model only to describe the decline of 1950–1954. Its maximum fell before the sheets begin, so its numbers are not given.
• Cycle 25 is still running. Its numbers will change as new observations come in; the fit is renewed by hand (install/cevrim_modeli.py, last on 2026-10-27).

Years usable / all sheets