paper

CHASM: coronal hole segmentation from SWPC drawings

2024 — 2025

Space Weather Prediction Center forecasters have outlined coronal hole boundaries on paper since the 1960s. Sixty years of expert judgment, sitting in an archive as scanned images — a labelled dataset nobody could actually use.

Forecaster pencil contour, scanned archive sheet.

CHASM digitizes them. The pipeline uses the Segment Anything Model to lift forecaster pencil contours off 1,111 synoptic drawings and reproject them into usable masks, with a manual review pass where the scan quality defeats it.

Method

With the dataset in hand, I trained a CHRONNOS/U-Net architecture to segment coronal holes directly from EUV imagery. It reaches 98% detection accuracy against the forecaster boundaries — which is to say it agrees with the pencil about as often as two forecasters agree with each other.

Results

Predicted mask vs. forecaster boundary, held-out synoptic map.

Detection accuracy across the 1,111-map evaluation set.

Agreement holds up across the eleven-year archive window rather than clustering in a few easy years, and most of the residual disagreement concentrates near the polar regions, where the pencil convention itself is least consistent between forecasters.

Paper