This repository contains the data for the article "Advection-free Convolutional Neural Network for Convective Rainfall Nowcasting" by Jenna Ritvanen, Bent Harnist, Miguel Aldana, Terhi Mäkinen, and Seppo Pulkkinen., submitted to IEEE JSTARS journal.

The model code used for generating the model checkpoints and nowcasts is available for the L-CNN model at https://doi.org/10.5281/zenodo.7118752 and for the RainNet model at https://doi.org/10.5281/zenodo.7118705.

composites

• composites: FMI radar composites used as input data for the article. The zip file contains the composites as gzip-compressed PGM images. The metadata of the composites is written in the first 29 lines of the file. The composites are given for the 100 days used in the article with a time interval of 5 minutes, resulting in 100 * 24 * 12 = 28 800 files, sorted in directories according to year, month and day. Note that no quality control has been performed on the composites. The compressed values can be transformed to radar reflectivity with dBZ = (data - 64.0) / 2.0. The projection of the composites is +proj=stere +a=6371288 +lon_0=25E +lat_0=90N +lat_ts=60 +x_0=380886.310 +y_0=3395677.920 +no_defs

models

Model checkpoints used to produce the nowcasts in the article.

• lcnn/epoch=6-step=95480.ckpt: L-CNN model
• rainnet/t11-rn-logcosh-lt30.ckpt: RainNet model

nowcasts

Nowcasts used to compute verification results in the article. The subimages were from the composites with bounding box [604, 1116, 125, 637], written as [x1, x2, y1, x2] that corresponds to image[x1:x2, y1:y2] in NumPy indexing.

• lcnn_diff_rmse_30lt_20062022_36.h5: L-CNN nowcasts
• p15-rn-logcosh-lt30.hdf5: RainNet nowcasts
• p25_extrapolation_lcnn_test_swap.hdf5: Extrapolation nowcasts
• p25_linda_lcnn_test_swap.hdf5: LINDA nowcasts
• test_obs_512.hdf5: Observations

verification_results

Verification statistic values in CSV files. The first row indicates leadtime index (i.e., leadtime = 5 min * value). The first column indicates statistic name and second the model.

• CONT.csv: continuous scores.
• CAT.csv: categorical scores. The statistic names follow the pattern _, e.g. CSI_10_0 for CSI at 10.0 threshold.
• FSS.csv: categorical scores. The statistic names follow the pattern __, e.g. FSS_16_10_0 for FSS at 16km scale at 10.0 threshold.