Skip to content

Latest commit

 

History

History
100 lines (61 loc) · 3.87 KB

README.md

File metadata and controls

100 lines (61 loc) · 3.87 KB

windflow-light

Perform optical flow inference on geostationary satellite images from a pretrained RAFT model.

windflow example

See python predict.py for a basic example.

Install

conda env create -f environment_cpu.yml

Usage

Usage for applying pretrained model to ECO1280 data

humidity.py

eco1280 example

gp_2016-06-01_00:00:00_P500_out.nc

of the form gp_yyyy-mm-dd_HH:MM:SS_P[pressurelvl]_out.nc

fields:

Field long name
gp_newP specific humidity
lev_p pressure level
lat_0 latitude
lon_0 longitude

UV Components:

uv_2016-06-01_00:00:00_P500_out.nc

of the form uv_yyyy-mm-dd_HH:MM:SS_P[pressure level]_out.nc

fields:

Field long name
ugrd_newP U component of wind
vgrd_newP V component of wind
lat_0 latitude
lon_0 longitude
lev_p pressure level

latitude, longitude, and pressure level are the same in both of these files.

  1. Read in UV Component file; extract lat, lon, u and v comp.
  2. Read in Humidity file; extract humidity data, scale or perform preprocessing if necessary ( This particular example scales by 25000 since the ECO1280 data was scaled by that amount. We want our humidity values to be in the 0 to 255 range)
  3. Load model checkpoint.
  4. perform inference on the model with the humidity data. The preprocessed humidity data is fed into the model to compute flows between the two time steps. This produces predictions for the u and v in units of pixels displacement
  5. Make plots of the truth and plots of the predictions. The V component is negated because of image orientation (see g5nr_eco1280_comparison.ipnyb)

scatter.py

scatter u

  1. Read in UV file and Humidity file
  2. Make windflow predictions on humidity data
  3. Create a mask for regions of interest (-30 to 30, or -60 to 60, -90 to 90)
  4. Convert Latitude / Longitude from degrees to radians and convert pixel displacement to meters/second

$$ \text{vcomp} = \text{vpixel}~ \cdot ~\frac{0.1 ~\text{deg}}{1 ~\text{pixel}} ~ \cdot ~ \frac{111 ~ \text{km}}{1 ~ \text{degree}} ~ \cdot ~ \frac{1000m}{\text{km}} ~ \cdot ~ \frac{1}{10800\text{sec}} $$

remember to negate vpixel before doing this.

convert pixel displacement to v component in units of m/s. Since these images are 3 hours apart, you divide by 3 * 60 * 60 = 10800 to get units of seconds.

$$ \text{ucomp} = \frac{1}{cos(lat)} ~ \cdot ~ \text{upixel}~ \cdot ~\frac{0.1 ~\text{deg}}{1 ~\text{pixel}} ~ \cdot ~ \frac{111 ~ \text{km}}{1 ~ \text{degree}} ~ \cdot ~ \frac{1000m}{\text{km}} ~ \cdot ~ \frac{1}{10800\text{sec}} $$

$\frac{1}{cos(lat)}$

  • This factor adjusts for the varying distance covered by a degree of longitude depending on the latitude. A degree of longitude represents a longer distance near the equator than it does closer to the poles. Since cos(lat) where latitude is in radians decreases as you move from the equator towards the poles, dividing by cos(lat) correctly adjusts the longitudinal distance covered per pixel.

There is density scatter, normal scatter, and plots with difference between real u/v and computed u/v.

contour.py

contour eco1280

Basically the same as humidity.py but instead of the original humidity data, this plots the wind speeds as a contour plot.

Citation

Vandal, T., Duffy, K., McCarty, W., Sewnath, A., & Nemani, R. (2022). Dense feature tracking of atmospheric winds with deep optical flow, Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining.

Acknowledgements

External packages and flownet code was used from: https://github.com/celynw/flownet2-pytorch/
Funded by NASA ROSES Earth Science Research from Geostationary Satellite Program (2020-2023)