Spatio-temporal prediction of missing temperature with stochastic Poisson equations

This is the implementation of the LC2019 team winning entry for the EVA 2019 data competition, as described in the paper.

Requirements

This project was implemented on Linux, on which it should work smoothly. It also works on Windows, without multiprocessing support (neither on WSL).

This project is mainly implemented with Python3. R is used to convert data. Programming environments of Python3 and R should be setup first. The Python dependencies are listed in requirements.txt. On a Ubuntu machine, R can be installed with

sudo apt install r-base

Usage

• Data preparation. Two data files can be downloaded via the link.

  • DATA_TRAINING.RData: for training and inference
  • TRUE_DATA_RANKING.RData: only for final evaluation.

  Put them in ./data/ folder. Convert them to .npy and prepare all necessary data.

  python data_conversion.py
  

• Build the utilities

  cd util
  python xmin_setup.py build_ext --inplace
  

• Compute x_min and split cross validation sets

  python compute_X_min.py
  python split_cross_validation.py
  

• Inference

  python PoissonTemperature.py 0 1000  # score: 3.61e-4
  

Bonus

The code here implements image cloning proposed in Poisson image editing. Try it with:

python poisson_image_editing.py

or .

Bibtex

@article{PoissonEVA2019,
    title = {Spatio-temporal prediction of missing temperature with stochastic {P}oisson equations},
    author = {Cheng, Dan and Liu, Zishun},
    year = 2021,
    journal = {Extremes},
    pages = {163--175},
    volume = {24},
    number = {1},
    issn = {1572-915X},
    url = {https://doi.org/10.1007/s10687-020-00397-w},
    doi = {10.1007/s10687-020-00397-w}
}

Related Projects

• BeatTheHeat
• BlackBox
• Multiscale