Park et al., 2019, ApJL, 884, L23, doi:http://doi.org/10.3847/2041-8213/ab46bb
Title: Generation of Solar UV and EUV Images from SDO/HMI Magnetograms by Deep Learning
For the UV and EUV images, we use SDO/AIA 9 passbands images.
For the solar magnetograms, we use SDO/HMI Line-of-sight magnetograms.
Let,
C(f, k, s) denotes as 2D Convolution layer with f filters, filter size of k, stride of s,
CT(f, k, s) as 2D Convolution-Transpose layer with f filters, filter size of k, stride of s,
B as Batch-Normalization layer,
R as ReLU activation layer,
L as Leaky-ReLU activation layer with slope 0.2,
T as Tanh activation layer,
S as Sigmoid activation layer,
and D as Dropout layer with rate 0.5.
We can select the size of the receptive field of the discriminator.
- 1x1 discriminator
C(64,1,1)-L-C(128,1,1)-B-L-C(1,1,1)-S
- 16x16 discriminator
C(64,4,2)-L-C(128,4,1)-B-L-C(1,4,1)-S
- 34x34 discriminator
C(64,4,2)-L-C(128,4,2)-B-L-C(256,4,1)-B-L-C(1,4,1)-S
- 70x70 discriminator
C(64,4,2)-L-C(128,4,2)-B-L-C(256,4,2)-B-L-C(512,4,1)-B-L-C(1,4,1)-S
- 142x142 discriminator
C(64,4,2)-L-C(128,4,2)-B-L-C(256,4,2)-B-L-C(512,4,2)-B-L-C(512,4,1)-B-L-C(1,4,1)-S
- 286x286 discriminator
C(64,4,2)-L-C(128,4,2)-B-L-C(256,4,2)-B-L-C(512,4,2)-B-L-C(512,4,2)-B-L-C(512,4,1)-B-L-C(1,4,1)-S
The generator network is consist of the encoder and the decoder
- C(64,4,2)-L
- C(128,4,2)-B-L
- C(256,4,2)-B-L
- C(512,4,2)-B-L
- C(512,4,2)-B-L
- C(512,4,2)-B-L
- C(512,4,2)-B-L
- C(512,4,2)-B-L
- C(512,4,2)-B-L
- C(512,4,2)-R
- CT(512,4,2)-B-D-R
- CT(512,4,2)-B-D-R
- CT(512,4,2)-B-D-R
- CT(512,4,2)-B-R
- CT(512,4,2)-B-R
- CT(512,4,2)-B-R
- CT(256,4,2)-B-R
- CT(128,4,2)-B-R
- CT(64,4,2)-B-R
- CT(1,4,2)-S
The generator network has skip-connections between i-th layers of the encoder and (10-i)-th layers of the decoder like the U-Net architecture.
- encoder 1st layer - decoder 9th layer
- encoder 2nd layer - decoder 8th layer
- encoder 3rd layer - decoder 7th layer
- encoder 4th layer - decoder 6th layer
- encoder 5th layer - decoder 5th layer
- encoder 6th layer - decoder 4th layer
- encoder 7th layer - decoder 3rd layer
- encoder 8th layer - decoder 2nd layer
- encoder 9th layer - decoder 1st layer
Keras: Change some parameters in option.py (about your environments) and run train.py
TensorFlow2: Change some parameters in solar_generation_tf2.ipynb and run the notebook