search

pySTEPS / pysteps

Python framework for short-term ensemble prediction systems.
https://pysteps.github.io/
BSD 3-Clause "New" or "Revised" License
453 stars 165 forks source link

Which flow is better? #87

Closed Fangyh09 closed 5 years ago

Fangyh09 commented 5 years ago

I found the procedure to compute optical flow is slow. And tried another way(neural network) to compute it. Which one looks better and reasonable? LK method or the new method? The results are: ------------------------------This repo LK method------------------------------ Time=t0: o_v2 Time=t1: o_v11111

------------------------------new method------------------------------ Time=t0: n_v1 Time=t1: n_v2

dnerini commented 5 years ago

Hi @Fangyh09 Interesting contribution! One way to compare optical flow estimations consists in computing the forecast error for both methods and see which one provides the best skill. See for example Section 5.1 in the pysteps reference publication, here: https://www.geosci-model-dev-discuss.net/gmd-2019-94/

Hope this helps.

Kind regards Daniele

Fangyh09 commented 5 years ago

@dnerini Thanks, I found ROC metric in https://pysteps.readthedocs.io/en/latest/auto_examples/plot_ensemble_verification.html#sphx-glr-auto-examples-plot-ensemble-verification-py.

dnerini commented 5 years ago

For example, yes, although the scores in that tutorial are for ensemble (probabilistic) forecasts, while you will be looking at deterministic predictions. You can therefore try to use deterministic continuous error metrics such as the mean absolute error and correlation coefficient, or categorical scores such as the critical success index (CSI). Those are the most commonly used scores. Alternatively, you can try spatial scores, which aim at providing the forecast skill for a range of spatial scales. You can find an example in this tutorial.

Again, I recommend you to have a look at the pysteps paper, where we have compared three different optical flow techniques (Section 5.1).

Fangyh09 commented 5 years ago

@dnerini Thanks.