Deep learning-based image-to-image translation from SDO/HMI magnetograms to SDO/AIA UV and EUV images.
Title: Generation of Solar UV and EUV Images from SDO/HMI Magnetograms by Deep Learning
Authors: Eunsu Park, Yong-Jae Moon, Jin-Yi Lee, Rok-Soon Kim, Harim Lee, Daye Lim, Gyungin Shin, Taeyoung Kim
Journal: The Astrophysical Journal Letters, 884:L23 (8pp), 2019 October 10
Code: GitHub Repository
This method applies deep learning to translate solar magnetograms into UV and EUV images. Two CNN models are compared: one using only L1 loss (Model A), and another using L1 + cGAN loss (Model B, Pix2Pix style).
| Type | Description |
|---|---|
| Input | SDO/HMI LOS magnetogram (1 channel) |
| Output | SDO/AIA UV/EUV image (1 channel per passband) |
| Passband (Å) | Temperature (K) | Region |
|---|---|---|
| 1700 | 4,500 | Temperature minimum, photosphere |
| 1600 | 10,000 | Transition region, upper photosphere |
| 304 | 50,000 | Chromosphere, transition region |
| 171 | 600,000 | Quiet corona, upper transition region |
| 193 | 1,000,000 | Corona, hot flare plasma |
| 211 | 2,000,000 | Active region corona |
| 335 | 2,500,000 | Active region corona |
| 94 | 6,000,000 | Flaring corona |
| 131 | 10,000,000 | Transition region, flaring corona |
The model is based on Pix2Pix (Isola et al., 2016) architecture with a U-Net Generator and PatchGAN Discriminator.
Input: 1024 × 1024 × 1 (magnetogram)
│
▼
┌─────────────────────────────────────────────────────────────────┐
│ ENCODER │
├─────────────────────────────────────────────────────────────────┤
│ Conv(1→64, k4s2) ─────────────────────────────────────┐ │
│ │ │ skip │
│ ▼ │ │
│ Conv(64→128, k4s2) → BN → LeakyReLU ──────────┐ │ │
│ │ │ skip │ │
│ ▼ │ │ │
│ Conv(128→256, k4s2) → BN → LeakyReLU ─────┐ │ │ │
│ │ │ │ │ │
│ ▼ │ │ │ │
│ Conv(256→512, k4s2) → BN → LeakyReLU ──┐ │ │ │ │
│ │ │ │ │ │ │
│ ▼ │ │ │ │ │
│ Conv(512→512, k4s2) → BN → LeakyReLU ┐ │ │ │ │ │
│ │ │ │ │ │ │ │
│ ▼ │ │ │ │ │ │
│ Conv(512→512, k4s2) → BN → LeakyReLU │ │ │ │ │ │
│ │ │ │ │ │ │ │
│ ▼ │ │ │ │ │ │
│ Conv(512→512, k4s2) → BN → LeakyReLU │ │ │ │ │ │
│ │ │ │ │ │ │ │
│ ▼ │ │ │ │ │ │
│ Conv(512→512, k4s2) → BN → ReLU │ │ │ │ │ │
│ │ (Bottleneck) │ │ │ │ │ │
└──────┼────────────────────────────────┼─┼──┼───┼───────┼────────┘
│ │ │ │ │ │
▼ ▼ ▼ ▼ ▼ ▼
┌─────────────────────────────────────────────────────────────────┐
│ DECODER │
├─────────────────────────────────────────────────────────────────┤
│ ConvT(512→512, k4s2) → BN → Dropout → ReLU + Skip │
│ │ │
│ ▼ │
│ ConvT(1024→512, k4s2) → BN → Dropout → ReLU + Skip │
│ │ │
│ ▼ │
│ ConvT(1024→512, k4s2) → BN → Dropout → ReLU + Skip │
│ │ │
│ ▼ │
│ ConvT(1024→512, k4s2) → BN → ReLU + Skip │
│ │ │
│ ▼ │
│ ConvT(1024→256, k4s2) → BN → ReLU + Skip │
│ │ │
│ ▼ │
│ ConvT(512→128, k4s2) → BN → ReLU + Skip │
│ │ │
│ ▼ │
│ ConvT(256→64, k4s2) → BN → ReLU + Skip │
│ │ │
│ ▼ │
│ ConvT(128→1, k4s2) → Tanh │
│ │ │
└──────┼───────────────────────────────────────────────────────────┘
│
▼
Output: 1024 × 1024 × 1 (UV/EUV image)
Input: 1024 × 1024 × 2 (magnetogram + image, concatenated)
│
▼
Conv(2→64, k4s2) → LeakyReLU
│
▼
Conv(64→128, k4s2) → BN → LeakyReLU
│
▼
Conv(128→256, k4s2) → BN → LeakyReLU
│
▼
Conv(256→512, k4s1) → BN → LeakyReLU
│
▼
Conv(512→1, k4s1) → Sigmoid
│
▼
Output: 62 × 62 × 1 (patch probabilities)
| Model | Loss Function | Characteristics |
|---|---|---|
| Model A | L1 only | Better metrics (CC, RE, PPE10), more blurred |
| Model B | L1 + cGAN | More realistic/sharper images, slightly lower metrics |
L1 Loss:
L1 = E[||y - G(x)||₁]
cGAN Loss:
L_cGAN = E[log D(x, y)] + E[log(1 - D(x, G(x)))]
Combined Loss (Model B):
G* = arg min_G max_D L_cGAN(G, D) + λ·L1(G)
where λ = 100
| Parameter | Value |
|---|---|
| Optimizer | Adam |
| Learning Rate | 2 × 10⁻⁴ |
| β₁ | 0.5 |
| β₂ | 0.999 |
| Iterations | 500,000 |
| Image Size | 1024 × 1024 |
| Conv Weight Init | Normal(0.0, 0.02) |
| BatchNorm Weight Init | Normal(1.0, 0.02) |
- Source: SDO/HMI LOS magnetograms and SDO/AIA 9-passband images
- Period: 2011-2017
- Cadence: 6 hours (4 pairs per day)
- Training: 8,544 pairs (2011-2016)
- Validation: 714 pairs (2017 Jan-Jun)
- Test: 727 pairs (2017 Jul-Dec)
- Preprocessing: Level 1.5 calibration, exposure time normalization, degradation correction
- AIA images: DN/s → log scale (0-14) → rescale to (-1, 1)
- HMI magnetograms: Rescale to (-1, 1)
| Passband | CC (Model A) | CC (Model B) | RE (Model A) | RE (Model B) |
|---|---|---|---|---|
| 1700 Å | 0.97 | 0.95 | 0.01 | 0.02 |
| 1600 Å | 0.94 | 0.92 | 0.11 | 0.11 |
| 304 Å | 0.84 | 0.83 | -0.18 | -0.17 |
| 171 Å | 0.69 | 0.66 | -0.05 | -0.04 |
| 193 Å | 0.78 | 0.74 | 0.07 | 0.07 |
| 211 Å | 0.86 | 0.78 | 0.08 | 0.08 |
| 335 Å | 0.86 | 0.85 | -0.03 | 0.04 |
| 94 Å | 0.79 | 0.75 | -0.03 | -0.02 |
| 131 Å | 0.81 | 0.78 | -0.04 | -0.03 |
| Average | 0.84 | 0.83 | 0.07 | 0.06 |
- Python 3.6+
- PyTorch 1.0+ (or TensorFlow/Keras as in original)
- NumPy
- SunPy
- SolarSoft (for data preprocessing)
import torch
from networks import Generator, Discriminator
# Initialize models
generator = Generator(in_channels=1, out_channels=1)
discriminator = Discriminator(in_channels=2) # magnetogram + image
# Input: (batch, 1, 1024, 1024) - HMI magnetogram normalized to [-1, 1]
magnetogram = torch.randn(1, 1, 1024, 1024)
# Output: (batch, 1, 1024, 1024) - Generated AIA image in [-1, 1]
generated_euv = generator(magnetogram)Title: Selection of Three (Extreme)Ultraviolet Channels for Solar Satellite Missions by Deep Learning
Authors: Daye Lim, Yong-Jae Moon, Eunsu Park, Jin-Yi Lee
Journal: The Astrophysical Journal Letters, 915:L31 (9pp), 2021 July 10
Type: Follow-up Study (Optimal Channel Selection for UV/EUV Translation)
Summary: This paper compares 170 deep learning models for image translation among SDO/AIA 9 channels to find optimal channel combinations:
- 170 Models Developed: 72 single-channel, 56 double-channel, 42 triple-channel input models
- Single-channel Best: 131 Å (average CC = 0.84)
- Double-channel Best: 131 + 1600 Å (average CC = 0.95)
- Triple-channel Best: 131 + 1600 + 304 Å (average CC = 0.97, minimum CC = 0.96)
- Physical Interpretation: The three channels represent coronal (131 Å), upper photospheric (1600 Å), and chromospheric (304 Å) layers
- Application: Results can guide UV/EUV channel selection for future solar satellite missions
@article{Lim_2021,
title={Selection of Three (Extreme)Ultraviolet Channels for Solar Satellite Missions by Deep Learning},
author={Lim, Daye and Moon, Yong-Jae and Park, Eunsu and Lee, Jin-Yi},
journal={The Astrophysical Journal Letters},
volume={915},
number={2},
pages={L31},
year={2021},
month={jul},
publisher={The American Astronomical Society},
doi={10.3847/2041-8213/ac0d54}
}@article{Park_2019,
title={Generation of Solar UV and EUV Images from SDO/HMI Magnetograms by Deep Learning},
author={Park, Eunsu and Moon, Yong-Jae and Lee, Jin-Yi and Kim, Rok-Soon and Lee, Harim and Lim, Daye and Shin, Gyungin and Kim, Taeyoung},
journal={The Astrophysical Journal Letters},
volume={884},
number={2},
pages={L23},
year={2019},
month={oct},
publisher={The American Astronomical Society},
doi={10.3847/2041-8213/ab46bb}
}MIT License