Machine learning for uncertainty quantification in climate precipitation prediction over South America
This repository contains the reference code for "Machine learning for uncertainty quantification in climate precipitation prediction over South America".
- LightGBM (3.x+)
- Optuna (3.x+)
- Pandas
- PyArrow
- SciTools Iris
To install all dependencies using Conda:
conda env create -f environment.ymlThis code was tested on Anaconda 2022.11 Pop!_OS 22.04.
The authors use GPCP (Global Precipitation Climate Project) v2, converted from NetCDF to Apache Feather. The dataset can be downloaded here.
In the conda environment, run python train.py. The training set are the years 1980-2017 with 75%/25% training split.
In the conda environment, run python analysis.py to generate the resulting maps. Place agcm.nc in the root directory, it contains the INPE/AGCM data.
The testing set are the years 2018 and 2019. Maps are generated for selected months (mid-season) of 2019.
Pre-trained models can be downloaded here. Place them and the Optuna databases inside the models folder and they will be used instead of training.
The figure below shows the precipitation maps along with estimated uncertainties for 4 selected months.
V. S. Monego, J. A. Anochi, H. F. de Campos Velho, South America seasonal precipitation prediction by gradient-boosting machine-learning approach, Atmosphere 13 (2). doi:10.3390/atmos13020243. URL https://www.mdpi.com/2073-4433/13/2/243
J. A. Anochi, R. Hern ́andez Torres, H. F. Campos Velho, Climate Precipitation Prediction with Uncertainty Quantification by Self-configuring Neural Network (International Symposium on Uncertainty Quantification and Stochastic Modelling – 2020), Vol. 2, Ch. 18.
R. F. Adler, G. J. Huffman, A. Chang, R. Ferraro, P.-P. Xie, J. Janowiak, B. Rudolf, U. Schneider, S. Curtis, D. Bolvin, et al., The version-2 global precipitation climatology project (gpcp) monthly precipitation analysis (1979–present), Journal of hydrometeorology 4 (6) (2003) 1147–1167.