PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-Modal Solar Image Synthesis: Adapting the Surya Foundation Model from He I 10830 {\AA} to EUV Translation and Coronal Hole Segmentation

Marco Marena, Andrés Muñoz Jaramillo, Qin Li, Haodi Jiang, Jinghao Cao, Wen He, Ziyang Zhang, Chenxi Yuan, Chao Wang, Haimin Wang, Bo Shen

arXiv:2610.04553v1astro-ph.SRastro-ph.IMcs.AIcs.CV

TL;DR

The paper asks whether the longer He I observational record can support EUV-like views of solar morphology before modern EUV imaging. It adapts Surya to predict AIA channels and CH probability maps from helium observations, with strong held-out global correlations and channel-dependent structural agreement.

  • Problem

    The shorter record of direct EUV imaging limits consistent comparisons of broad solar morphology across eras, motivating tests of what EUV and CH structure He I observations can predict.

    Source: §I · p. 1

  • Method

    The paper adapts Surya with a convolutional input adapter, LoRA backbone updates, and dedicated decoders to predict AIA intensities and CH probabilities from He I observations.

    Source: §VII · p. 9

  • Results

    Held-out dedicated models reached CCs of 0.8196, 0.8885, and 0.8398 for AIA 94, 193, and 304, respectively; the CH model achieved IoU = 0.4360.

    Source: §VII · p. 9

  • Takeaways & Limitations

    The results support helium-conditioned EUV morphological proxies for historical reconstruction, not replacement of archival observations or unique inversion of atmospheric state.

    Source: §VI · p. 9

    The historical EIT comparison covers only three selected dates on a different grid and has instrument-response, resolution, and quality-flag differences.

    Source: §VI · p. 9

Abstract

from arXiv · show

The long observational record of He I 10830 {\AA} offers a means to investigate solar morphology before modern extreme-ultraviolet (EUV) imaging. We adapt the Surya solar foundation model to predict Solar Dynamics Observatory/Atmospheric Imaging Assembly (SDO/AIA) 94, 193, and 304 {\AA} images and a coronal hole (CH) probability map from full-disk helium observations. A convolutional input adapter, low-rank backbone updates, and dedicated output decoders learn from temporally paired, geometrically registered observations, with Spatial Possibilistic Clustering Algorithm (SPOCA) catalog polygons providing CH supervision. On observations held out from downstream fine-tuning, the selected dedicated models achieve disk-restricted correlation coefficients (CCs) of 0.8196, 0.8885, and 0.8398 for the three AIA channels, respectively; the CH model achieves an intersection over union (IoU) of 0.4360. The predictions recover broad solar structure, although local agreement varies substantially by channel. An optional residual refiner addresses spatial detail, and its application on pre-SDO dates improves the correlation of synthetic AIA 304 with Solar and Heliospheric Observatory/Extreme-ultraviolet Imaging Telescope (SOHO/EIT) 304 references from 0.6511 to 0.7085. Together, these results support the feasibility of helium-conditioned EUV morphological proxies and motivate their use in historical reconstruction, within the scope of the downstream test and cross-instrument evaluation.

I. INTRODUCTION

The paper asks whether full-disk He I 10830 Å observations can support reconstruction of multiple EUV morphologies and coronal-hole maps, addressing the shorter observational record of EUV imaging. It adapts Surya for separate EUV-intensity and CH-probability outputs and frames historical inference as a use of the resulting models.

I. INTRODUCTION

  • The study asks how much morphology in several EUV channels and CH maps can be predicted from full-disk He I 10830 Å observations.The motivation is to compare broad bright and dark structures across eras using a consistent EUV representation, given the longer record of chromospheric proxies.Source: §I · p. 1
  • He I morphology is a proxy for large-scale coronal structure, but different atmospheric states can produce similar helium signatures.The triplet forms in the upper chromosphere, with its population influenced by coronal and transition-region radiation and potentially by collisional excitation.Source: §I · p. 1
  • The paper treats CH segmentation as probabilistic rather than intensity regression because dark EUV signatures do not uniquely define CH boundaries.SPoCA supplies algorithmic labels using fuzzy or possibilistic clustering with spatial and temporal processing.Source: §I · p. 1
  • The paper extends prior He I-to-AIA translation work by adapting an existing solar foundation model to several distinct target products.Surya's pretrained representation is adapted to a single ground-based input; a separate residual refiner investigates local corrections to broad-structure outputs.Source: §I · p. 2
  • The work pairs and registers SOLIS He I, AIA 94/193/304 Å, and SPoCA CH records, then adapts Surya with a convolutional input adapter, LoRA, and dedicated decoders.It also tunes output-specific configurations and evaluates historical inference against independent EIT references on three dates.Source: §I · p. 2

II. DATASET AND PHYSICAL ALIGNMENT

The dataset pairs full-disk SOLIS He I observations with AIA intensity targets and SPoCA coronal-hole labels on a common grid. Temporal matching and month-based partitions define the quality-controlled cohorts used for training, validation, and downstream testing.

A. Observations and scientific targets

  • The SOLIS input is a full-disk He I 1083.0 nm scalar product, reprojected to a 4096 × 4096 solar-north-up grid; targets are AIA 94, 193, and 304 Å scalar arrays.The sampling grid matches Surya and the target grid but does not increase SOLIS optical resolution.Source: A. Observations and scientific targets · p. 2
  • SPoCA event polygons are associated within 120 minutes, rasterized onto the AIA grid, and bilinearly reprojected into soft CH boundary targets.These are operational algorithmic labels, not exhaustive expert annotations; early-year event provenance and missing-versus-empty masks require audit.Source: A. Observations and scientific targets · p. 2

B. Temporal association and data partitioning

  • Each SOLIS observation is matched to the nearest admissible AIA image per channel within 15 minutes, while SPoCA associations use a separate 120-minute tolerance.The manifest records timestamps, offsets, paths, and quality status; it defines the experiment cohort.Source: B. Temporal association and data… · p. 2
  • The common-grid AIA 193/304 cohort uses January–October for training, November for validation, and December for downstream testing, with 57 test observations on 54 dates.The test observations are excluded from downstream fine-tuning and validation-based selection, but there is no buffer at month boundaries.Source: B. Temporal association and data… · pp. 2–3

C. World coordinate system registration and preprocessing

The preprocessing aligns modalities through solar WCS metadata and evaluates them on a disk-centered support. Per-image percentile normalization emphasizes morphology and contrast rather than absolute radiometric calibration.

C. World coordinate system registration and preprocessing

  • Images are registered from solar WCS metadata onto a centered, solar-north-up helioprojective grid rather than aligned by resizing array indices.The catalog mask receives the same geometric transform, retaining fractional boundary values as soft targets.Source: C. World coordinate system… · p. 3
  • Evaluation uses a circular support with radius 0.48N, stored with each sample, to restrict scoring to the solar disk.The native-resolution experiments retain 4096 × 4096 pixels.Source: C. World coordinate system… · p. 3
  • Per-image normalization maps intensities between the first and 99th percentiles to [0, 1], clipping values outside that range.Metrics therefore characterize normalized morphology and contrast, not absolute radiometric calibration; clipping also removes information in the brightest tail.Source: C. World coordinate system… · p. 3

III. SURYA ADAPTATION

The adaptation converts a single He I image into Surya-compatible inputs, updates selected backbone projections with LoRA, and trains separate decoders for each output. The dedicated models share a pretrained starting point but learn independently.

III. SURYA ADAPTATION

  • Four dedicated models start from the same pretrained Surya checkpoint but learn separate input adapters, LoRA parameters, and single-output decoders.The four products are AIA 94, AIA 193, AIA 304, and CH probability; a separate exploratory configuration predicts all four jointly.Source: §III · p. 3
  • A convolutional adapter maps the single-channel He I input to 13 channels before the pretrained backbone processes it.The front end uses two 3 × 3 convolutional blocks and a final 1 × 1 projection.Source: §III · p. 3
  • The adapted He I image is repeated across Surya's two expected time slots, preserving its temporal-embedding interface without adding temporal information.Selected backbone projections are fine-tuned with low-rank updates while pretrained base matrices remain frozen.Source: §III · p. 3
  • Dedicated one-channel decoders reconstruct image patches; only the selected AIA 304 configuration adds a token MLP, while the CH decoder converts logits to probabilities with sigmoid.The input adapter, selected decoder, and low-rank parameters are trained together.Source: §III · p. 4

IV. EXPERIMENTAL DESIGN

EUV intensity training uses losses restricted to the solar disk, with optional spatial or structural terms, while coronal-hole prediction combines binary cross-entropy and soft Dice. Validation performance guides configuration selection, and the reported single-image batches do not reduce the contribution of less-synchronous pairs through sample weighting.

A. Training objective and model selection

  • AIA reconstruction loss uses only pixels inside the solar-disk mask, with a base L1 objective and optional gradient-consistency or SSIM structural terms.The selected (βc, γc) values are (0.1, 0), (0.1, 0.1), and (0, 0) for AIA 94, 193, and 304, respectively.Source: A. Training objective and model… · p. 4
  • With batch size one, normalized weighted-mean aggregation cancels the sample weight, so less-synchronous pairs are not downweighted in reported batches.Source: A. Training objective and model… · p. 4
  • CH prediction combines binary cross-entropy on logits with soft Dice loss to penalize pixelwise classification errors and region-overlap disagreement.The reported CH loss uses wBCE = 2 and wDice = 1.Source: A. Training objective and model… · p. 4
  • Validation MAE selects AIA configurations and validation Dice selects CH configurations; the retained training checkpoint and the metric used to compare runs are reported separately.The CH setup uses λCH = 5, wBCE = 2, and wDice = 1.Source: A. Training objective and model… · p. 4

B. Experimental controls and reproducibility

Experimental controls keep rank-sweep comparisons fixed across key settings, but the study does not isolate the contribution of Surya pretraining. Joint-versus-dedicated results are descriptive because their optimization objectives differ.

B. Experimental controls and reproducibility

  • Each rank-sweep row changes rank while retaining its manifest, grid, normalization, loss, and decoder; separate loss or decoder follow-ups are not treated as pure rank effects.Source: B. Experimental controls and… · p. 4
  • No matched random-initialization or small-network baseline estimates how much accuracy is specifically due to Surya pretraining.Source: B. Experimental controls and… · p. 4
  • Joint-versus-dedicated comparisons are descriptive, older full-weight and frozen-backbone runs are excluded from native-resolution comparisons, and headline AIA scores concern normalized scalar arrays rather than RGB renderings.Source: B. Experimental controls and… · p. 4

C. Active-region-aware residual refinement

The residual refinement stage targets compact bright cores and fine active-region structure that dedicated Surya–LoRA models can suppress. It adds a local correction to the global prediction, using patch-based training and active-region-focused sampling.

C. Active-region-aware residual refinement

  • Dedicated Surya–LoRA predictions recover broad EUV-disk organization but can suppress compact bright cores and fine active-region structure.Source: C. Active-region-aware residual… · p. 4
  • Separate residual attention U-Nets predict corrections from normalized He I input and frozen Surya predictions, with patch training and active-region-focused sampling and loss weighting.Source: C. Active-region-aware residual… · p. 5
  • The residual formulation learns an additive adjustment rather than synthesizing the disk anew; conditioning on He I retains access to the observation, but does not guarantee physically correct sharpened features.Source: C. Active-region-aware residual… · p. 5
  • Refiners train only on paired SDO-era observations; HARP masks support active-region tile oversampling during training but are not inference inputs.Historical refinement uses He I and a synthetic AIA prediction, without EIT images or historical targets in training.Source: C. Active-region-aware residual… · p. 5

D. Evaluation protocol

The evaluation compares output-specific configurations on validation data, then tests retained checkpoints on a separate 57-observation December cohort. Results distinguish global correlation from local structure, with uncertainty estimated across test dates.

D. Evaluation protocol

  • Continuous outputs use disk-restricted image-level CC, MAE, PSNR, and SSIM, while CH overlap metrics threshold predictions and targets at 0.5.AIA summaries average 57 observations, and calendar-day-cluster bootstrap intervals avoid treating correlated pixels as independent samples.Source: D. Evaluation protocol · p. 5

V. RESULTS

  • Validation selects AIA checkpoints by MAE and CH by Dice before retained models are evaluated on a separate 57-observation December cohort.The test outputs are on a 4096 × 4096 grid; Fig. 2’s archived AIA 94/304 examples are from earlier checkpoints, not the exact rank-64 models in Table IV.Source: §V · p. 5, Figure 2 · p. 6

A. Output-specific hyperparameter tuning

  • Under fixed linear-L1 comparisons, validation MAE is lowest at rank 64 for AIA 94/304 and rank 32 for AIA 193; separate loss and decoder follow-ups further tune outputs.The displayed CH rank sweep favors rank 16, but its provenance differs from the earlier dedicated CH checkpoint reported in Table IV.Source: A. Output-specific hyperparameter tuning · p. 5

B. Selected checkpoints and held-out performance

  • AIA 193 provides the strongest structural match, while AIA 94’s low SSIM indicates local discrepancies; bootstrap intervals capture date-sampling uncertainty, not label error or seed variability.The retained-checkpoint evaluation uses native-grid outputs restricted to the WCS solar disk.Source: B. Selected checkpoints and held-out… · pp. 5, 7, Table IV · p. 7

C. Spatial fidelity and the motivation for refinement

Direct predictions recover broad solar organization but retain local structural errors, motivating refinement. On selected pre-SDO dates, refined synthetic AIA 304 agrees more closely with EIT 304 references, which are cross-instrument comparisons rather than AIA ground truth.

C. Spatial fidelity and the motivation for refinement

  • AIA 193 performs strongly in both global correlation and local structural measures, while larger CC–SSIM gaps for AIA 94 and, to a lesser extent, AIA 304 mark remaining structural error.The metric gap does not establish a physical or architectural cause.Source: C. Spatial fidelity and the motivation… · p. 7
  • Across three illustrative held-out dates, predictions place prominent active regions and broad dark structures approximately correctly, but smooth fine emission and yield less precise CH boundaries.The illustrated dates were selected for display rather than sampled randomly.Source: C. Spatial fidelity and the motivation… · p. 7
  • Fig. 2 illustrates model-family behavior, not visual validation of Table IV’s exact AIA 94/304 rank-64 checkpoints, because its panels use earlier dedicated models.Source: C. Spatial fidelity and the motivation… · p. 7

D. Historical synthesis and cross-instrument refinement

  • Historical evaluation applies retained models to 27 pre-2010 records with EIT observations within 15 minutes of He I, without EIT fitting, histogram matching, or EIT-derived supervision.EIT 195 and AIA 193 are not the same passband, and EIT is an independent reference rather than AIA ground truth.Source: D. Historical synthesis and… · p. 7
  • Refinement raises synthetic AIA 304–EIT 304 CC from 0.6511 to 0.7085 and SSIM from 0.3610 to 0.4535 on a three-date, 1024 × 1024 disk evaluation.MAE decreases from 0.1032 to 0.0668 and PSNR increases from 16.985 to 20.597 dB; these are exploratory cross-instrument results, not native-grid AIA test metrics.Source: Table V · p. 7, D. Historical synthesis and… · p. 9

VI. DISCUSSION

Dedicated He I-conditioned models recover substantial morphology across EUV channels and coronal-hole maps, but agreement varies by channel and fine detail. Historical refinement improves agreement with EIT on three dates, while the limited tests constrain stronger physical or historical claims.

VI. DISCUSSION

  • Disk-restricted CCs are strongest for AIA 193, while lower AIA 94 SSIM shows that global correlation does not ensure recovery of fine hot-coronal texture.The wavelengths are distinct synthesis tasks, not one generic EUV representation.Source: §VI · p. 9
  • The trained workflow needs only a helium observation at inference to generate synthetic AIA-like morphological proxies for pre-AIA dates, not to replace archives or uniquely invert atmospheric state.The historical goal is a common representation for comparing broad solar structure across eras.Source: §VI · p. 9
  • CC improved from 0.6511 to 0.7085 after refinement in the three-date AIA 304–EIT 304 comparison, but this does not establish that added fine structures are physically correct.WCS registration and disk-restricted scoring are essential to this comparison.Source: §VI · p. 9
  • Validation-based rank selection should be output-specific: rank-128 follow-ups did not improve on rank 64, while AIA 193 favored rank 32 in its linear-head sweep.The refiner is a separate spatial-scale intervention, judged against unrefined predictions on the same observations with identical registration, normalization, and metric support.Source: §VI · p. 9
  • The 54-date December test, three-date EIT comparison, and unresolved data, label, and evaluation audits constrain claims about physical detail and broad historical generalization.The paper identifies these audits and controlled evaluations as prerequisites for stronger physical claims.Source: §VI · p. 9

VII. CONCLUSIONS

The paper adapts Surya to synthesize three AIA intensity proxies and coronal-hole probability maps from one He I observation. Held-out results show complementary EUV reconstructions and segmentation, while historical EIT comparisons improve with refinement on three dates. The products remain normalized morphological proxies rather than calibrated radiance or uniquely determined atmospheric states.

VII. CONCLUSIONS

  • A convolutional input adapter, LoRA updates, and dedicated decoders adapt Surya to synthesize AIA 94, 193, and 304 intensity proxies and CH probability maps from one He I observation.Validation tuning identifies different preferred configurations across channels rather than one rank and objective for every target.Source: §VII · p. 9
  • Disk-restricted CCs were 0.8196, 0.8885, and 0.8398 for AIA 94, 193, and 304, respectively, and the separately retained CH model reached F1 = 0.6073 and IoU = 0.4360.These held-out results support multiple complementary EUV reconstructions and an algorithmically supervised segmentation product.Source: §VII · p. 9
  • CC improved from 0.6511 to 0.7085 and SSIM from 0.3610 to 0.4535 after refinement in the three-date historical AIA 304–EIT 304 evaluation.The workflow can be applied to historical helium observations without a contemporary AIA image; the EIT references were not inference inputs.Source: §VII · pp. 9–10
  • The products are normalized morphological proxies, not calibrated radiance measurements or uniquely determined atmospheric states.The authors identify extended historical evaluation, matched full-test active-region assessment, and uncertainty quantification as next steps.Source: §VII · p. 10
Loading 2610.04553v1…