cnn-precip-predictor

This project aims to improve preiciptation forecast from GEFS with deep learning methods. With the capability of connecting spatial patterns to abstract concepts, CNN could potentially provide improvement to medium range regional forecast of Sacramento River Basin from the forecasts of a larger spatial region.

Data obtained from (https://psl.noaa.gov/forecasts/reforecast2/download.html). GEFS went through an update from v2 to v12 in Sept. 2020. GEFS folder contains preprocesing files. data/gefs-merge-two-files.py aggregates GEFS forecast to daily forecasts. data/to_tensor.py converts preicpitation data to tensors input. In addition to the preicipitation layer, two more layers of lon/lat features are included. (Currently this procedure is done by keras_experiment.py in Keras folder).

Keras contains CNN and MLP models to generate probablistic predictions for target region.

Main Run run_local.py to complete training of one model. Jupyter notebook with the name tuning helps monitor training quality. Run job_slurm_array.sh on a computing cluster to train models in batch. Run outcome_matrix.sh to obtain classifier outcomes on the cluster instead of on a PC.

Analysis evaluates model results with F1 score and ROC, and creates the benchmarks. Two benchmarks are available. The naive GEFS benchmark makes prediction by comparing the forecasted spatial average with the actual precipitation threshhold. The bias-corrected GEFS makes prediction essentially from ranking the forecasted spatial average.

Results gathers all .csv files resulted from the experiment, benchmark and analysis. In addition, a sample model and saliency file is provided.

• outcomes contains probabilistic predictions derived from model (1985-2019)

• f1 contains calculations for F1 score.

• confusion_matrix contains binary results from models and benchmarks

• 8_outcomes contains 222 contigency table comparing ERA5 ground truth, bias-corrected GEFS and model results.

Saliency allows daily/categorical investigation on which input "pixels" contribute to a positive/negative classification.

Figures contains codes to generate all figures in the figures.pdf, including visualization of analysis, saliency, reliability diagram, specific examples and visualization of filters and features maps.

Package requirements: xarray, keras, cartopy Tensorflow version: 2.3.0 Python 3.7.10

Reference: https://github.com/jdherman/cnn-streamflow-forecast