Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution

In this paper, we strive to construct a conditional diffusion model by incorporating the brightness temperature data from GridSat satellite observations. Additionally, during the sampling process, we will add guidance from the low-resolution ERA5 to generate high-quality high-resolution ERA5 maps.

Siwei Tu¹  Ben Fei^2,3,†  Weidong Yang^1,†  Fenghua Ling²  Hao Chen²  Zili Liu⁴  Kun Chen^1,2  Hang Fan^2,5  Wanli Ouyang^2,3  Lei Bai² 

¹Fudan University  ²Shanghai Artificial Intelligence Laboratory  ³Chinese University of Hong Kong  ⁴Beihang University  ⁵Nanjing University of Information Science and Technology 

Teaser of SGD:

Framework of SGD:

---

♦️ Training

The code is the March version.

Before starting the training, you can independently set the parameter values to be used, as shown below

MODEL_FLAGS="--image_size 512 --num_channels 64 --num_res_blocks 2"
DIFFUSION_FLAGS="--diffusion_steps 1000 --noise_schedule linear"
TRAIN_FLAGS="--lr 1e-4 --batch_size 1"

Use the image_train.py to start training.

python image_train.py --channel 4 --cond_channel 3 --size_x 720 --size_y 1440 --cond_x 2000 --cond_y 5143 $MODEL_FLAGS $DIFFUSION_FLAGS $TRAIN_FLAGS

🔥Sampling

Utilize the low-resolution ERA5 maps and w5k data to guide the sampling process.

python image_sample.py $MODEL_FLAGS $DIFFUSION_FLAGS

Utilize the w5k data to guide the sampling process by adjusting the loss function in image_sample_guide.py.

python image_sample_guide.py $MODEL_FLAGS $DIFFUSION_FLAGS

👏 Acknowledgement

Our paper is inspired by:

• https://generativediffusionprior.github.io/(the GDP repo)
• https://github.com/Stability-AI/stablediffusion.git(the Stable Diffusion repo)

Thanks for their awesome works!

If you have any inquiries, please feel free to consult via email [email protected] .