Machine Learning Based Identification of Solar Disk and Plages in Kodaikanal Solar Observatory Historical Suncharts

TL;DR

This paper addresses the challenge of extracting plages from historical KoSO hand-drawn suncharts by developing a two-stage CNN-based pipeline (U-Net) for disk detection and plage segmentation. Disk geometry is refined with Canny and Hough transforms, while plages are detected on 2048×2048 suncharts using a patch-based U-Net with a ResNet-34 encoder and CVAT-ground-truth masks, achieving robust performance ($IoU \approx 0.8$). The resulting plage masks enable time-latitude diagrams and a plage-area series that show strong agreement with Ca II K full-disk observations ($r = 0.80$, $\rho = 0.85$), allowing the construction of a composite dataset that fills historical data gaps. Limitations include non-detections in pre-1916 due to grid differences and 1990 annotation peculiarities, with future work to generalize to other features, harmonize training data, and publish the data for community use.

Abstract

Kodaikanal Solar Observatory (KoSO) is one of the oldest solar observatories, possessing an archive of multi-wavelength solar observations, including white light, Ca II K, and H-alpha images spanning over a century. In addition to these observations, KoSO has preserved hand-drawn suncharts (1904-2022), on which various solar features such as sunspots, plages, filaments, and prominences are marked on the Stonyhurst grid with distinct colour coding. In this study, we present the first comprehensive result that includes the entire data set from these suncharts using a supervised Machine Learning model called "Convolutional Neural Networks (CNNs)", firstly to identify the solar disks from the charts (1909-2007), secondly to identify the plages, spanning 9 solar cycles (1916-2007). We train the CNN with the manually identified solar disk and plage. We first detect the solar limb and the North-South line in the suncharts, which enables the extraction of disk centre coordinates, radius, and P-angle. Following that, we use a CNN similar architecture to achieve accurate image segmentation for the identification of plages. We compare plage areas derived from the suncharts with those obtained from Ca II K full-disk observations, and find good agreement that demonstrates the potential application of such an ML technique for historical data. The results of this study further demonstrate the potential application of sunchart data to fill the existing data gaps in the KoSO multi-wavelength observations and contribute toward constructing a composite series over the last century.

Figures (17)

• Figure 1: Temporal coverage of solar drawings from various observatories, showing the duration of their observations in years. The KoSO provides the longest continuous dataset (1904 -- 2022), followed by Meudon, MWO, Kanzelhöhe, and others.
• Figure 2: Representative suncharts produced at KoSO. (a) A sunchart from the KoSO dated January 5, 1958, showing detailed hand-drawn annotations of sunspots, plages, filaments, and prominences observed by the KoSO. (b) Another sunchart produced at KoSO dated 1958 February 06, showing solar features with different color as the original observation was made by a different observatory. Please note that suncharts are contrast-enhanced for better visualisation.
• Figure 3: Comparison of the annual number of Ca ii K images with the total number of suncharts and those containing plage markings, represented by the number of observing days. This highlights periods where suncharts, particularly those with plages (1955 -- 1988), can supplement or extend the Ca ii K observational record.
• Figure 4: (Left): U-Net architecture (adapted from ronneberger2015) used for segmenting plages. The architecture consists of a contracting path (left) for feature extraction and an expansive path (right) for precise localization. Key components include 3$\times$3 convolutions with ReLU nonlinearity (blue arrows), 2$\times$2 max pooling (dark red arrows), 2$\times$2 up-convolutions (light blue arrows), and skip connections for combining high-resolution features from the contracting path with the upsampled output. (Right): Flowchart showing the steps for disk and plage detection in suncharts.
• Figure 5: (a) A representative training image with an annotated training mask (red). (b) The corresponding binary mask used for training the disk detection model.