Sato and Ise (submitted) Open Data

______________________________________________________
Sato and Ise (2021) Open Data
Author: Hisashi SATO (JAMSTEC) hsatoscb_(at)_gmail.com
Date: 22 Feb 2022
URLs:
https://ebcrpa.jamstec.go.jp/~hsato/Sato_and_Ise_2020

Corresponding publication: 
H. Sato and T. Ise (2021) Predicting global terrestrial biomes with the LeNet convolutional neural network. Geoscientific Model Development 2022 Vol. 15 Issue 7 Pages 3121-3132. DOI: 10.5194/gmd-15-3121-2022
______________________________________________________
1. Folder "1.VCDs"
It contains Visualized Climate Environments (VCEs) for training and testing the Convolutional-Neural-Network (CNN) model. Names of compressed files (*.tar.gz) correspond to experiments. Each compressed file contains 4 to 16 folders.
Naming rule of folders:
The strings "CRU," "NCEP," "HadGEM," and "Miroc" stand for that the files are made from CRU_TS4.0, NCEP/NCAR reanalysis, Had2GEM-ESM, and Miroc-ESM climate datasets, respectively. The strings "AMeans" and "MMeans", respectively, stand for the annual-mean and monthly-mean climates, which are represented by VCEs in the compressed file.
The string "EachYear" means that the compressed files contains VCEs of each year climate from 1971 to 1980, otherwise the compressed files contains VCEs of averaged climate over 10 years. The strings "hist", "RCP26", and "RCP85" mean that the compressed files contain VCEs of climate averaged over 1971-1980, 2091-2100@RCP2.6, and 2091-2100@RCP8.5, respectively.
MainSimulation_training.tar.gz
VCEs for training the CNN model for the main simulation and dependency test of climatic datasets for training and reconstructing performances.
MainSimulation.tar.gz
VCEs for testing the CNN model for the main simulation.
CombinationSelection_Amean.tar.gz
VCEs for an experiment to find optimal combination of climatic variables (annual means) represented by the VCEs. In the VCE of the RGB color tile, up to three climate variables can be represented by RGB channels. To find the optimal combination of climatic variables, we systematically evaluated the model performance of 14 combinations of climatic variable experiments for annual means. It's result is presented in the Table S3.
Extracting this compressed fire results in "CRU", "NCEP", "HadGEM", and "Miroc" folders. Each folder contains 14 subfolders named with sequential numbers, which corresponds to model numbers in the Table S3.
CombinationSelection_Mmean.tar.gz
Same as the CombinationSelection_Amean.tar.gz except made with monthly mean climates, and related information is available in the Table S4.
ScalerSelection.tar.gz
VCEs for evaluating the influences of different transformations of climatic variables on the resulting accuracy. It's result is presented in the Table S5.
Extracting this compressed fire results in "CRU", "NCEP", "HadGEM", and "Miroc" folders. Each folder contains 4 subfolders named with sequential numbers. "Scaler1" stands for log transformed, "Scaler2" stands for no transformed (linear), "Scaler3" stands for Sigmoid(gain=5) transformed, "Scaler4" stands for Sigmoid(gain=10) transformed.
ColorAssignExperiment.tar.gz
VCEs for evaluate the influences of assignment patterns of air temperature and precipitation to RGB color channels of the VCE. It's result is presented in the Table S6. Extracting this compressed fire results in "CRU", "NCEP", "HadGEM", and "Miroc" folders. Each folder contains 6 subfolders named with sequential numbers, which indicates model number in the Table S6.
AverageExtentExperiment.tar.gz
VCEs for sensitivity test, where training and test accuracies are compared among models that are trained by monthly climate averaged over three different periods: 10 years (1971-1980; Control), 20 years (1961-1980), and 30 years (1951-1980).
Individual VCE shows climatic condition of each half degree grid cell. According to the ISLSCP2 data, VCEs are classified by their their potential vegetation type of the grid. Number of deepest folder names correspond vegetation code of the ISLSCP2.
Following is the vegetation code.
  01: Tropical Evergreen Forest/Woodland
  02: Tropical Deciduous Forest/Woodland
  03: Temperate Broadleaf Evergreen Forest/ Woodland
  04: Temperate Needleleaf Evergreen Forest/Woodland
  05: Temperate Deciduous Forest/Woodland
  06: Boreal Evergreen Forest/Woodland
  07: Boreal Deciduous Forest/Woodland
  08: Evergreen/Deciduous Mixed Forest
  09: Savanna
  10: Grassland/Steppe
  11: Dense Shrubland
  12: Open Shrubland
  13: Tundra
  14: Desert
  15: Polar Desert/Rock/Ice
The naming rule for VCEs of 10 years average climate is following:
Latitude number + "_" + Longitude number + ".png"
The naming rule for VCEs of each year's climate is following:
Latitude number + "_" + Longitude number + "_" + Year of climate + ".png"
Here, latitude number ranges from 001 to 360, starting from north-latitude-90 to southward with half degree interval. The longitude number ranges from 001 to 720, starting from west-longitude-180 to eastward with half degree interval. 
On top of this subfolder, there is PicList.zip, which is a compressed file of PicList.txt. This text file is required to classify VCEs with the trained model.
_____________________________________________
2. Folder "2.Result"
It contains copies of the image classification results on the Digits screen. 
Subfolder names correspond to experiment names. Each subfolder contains sub sub-folders "DigitsOutput," "ConfusionMatrix," and "ReconstructedBiomeMap." Naming rules for files are basically same as those of VCEs.
_____________________________________________
3. Folder "3.LearningCurves"
It contains screen capture of Digits output, showing learning curve of CNN models. CRU climate data. This folder contains 5 subfolders. Naming rules of these subfolders are same as those of VCEs.
_____________________________________________
4. Folder "5.FigureMaterials" 
It contains data and codes (in R) for drawing the figures in the manuscript."
_____________________________________________________________