Climate-Land Cover data used by DeepEcoClimate
Description
Sourced and processed from - Climate normals: CRU TS v4.06 - Land cover: MCD12C1 - Elevation: GMTED2010 ------------------------------------- climate_data_land.h5 contains 7 datasets with gzip compression 'indices' - shape (66501, 2) - latitude, longitude for all land pixels 'elev' - shape (66501, ) - average elevation (in metre) for each land pixel 'tmp' - shape (66501, 124, 12) - average daily mean temperature of each month during 1901-2024 for each land pixel 'pre' - shape (66501, 124, 12) - precipitation of each month during 1901-2024 for each land pixel 'pet' - shape (66501, 124, 12) - potential evapotranspiration (PET) of each month during 1901-2024 for each land pixel 'tmn' - shape (66501, 124, 12) - average daily minimum temperature of each month during 1901-2024 for each land pixel 'tmx' - shape (66501, 124, 12) - average daily maximum temperature of each month during 1901-2024 for each land pixel Temperatures are in Celcius, precipitation and PET are in millimeters 'indices', 'elev', 'pet' are not used in the climate classification models, only for plotting ------------------------------- climate_variables.h5 contains 1 dataset with gzip compression 'res' - shape (66502, 124, 10) res[-1, :, :] is the global average Ten climate variables during 1901-2024 for each land pixel "Coldest Month Mean Temperature": 0, "Hottest Month Mean Temperature": 1, "Coldest Month Mean Daily Minimum": 2, "Hottest Month Mean Daily Maximum": 3, "Mean Annual Temperature": 4, "Wettest Month Precipitation": 5, "Driest Month Precipitation": 6, "Total Annual Precipitation": 7, "Thermal Index": 8, "Aridity Index": 9, Temperature in Celcius, precipitation in millimeters These data facilitates plotting in the ClimViz app ------------------------------------- data.mat contains 4 ndarrays 'inputs' - shape (1411486, 3, 12) - 1411486 observations, tmn, pre, tmx for 12 months 'targets' - shape (1411486, 14) - 1411486 observations, proportions for 14 land cover types 'features' - shape (1411486, 60) - 1411486 observations, 60 features derived from MATLAB-pretrained model 'new_centroid' - shape (26, 60) - 26 centroids of climate types in the feature space, manually merged from a (5, 4, 3) self-organizing map trained in MATLAB This file is used in TorchModel.py (in the GitHub repo) to migrate (and regularize) the model from MATLAB to PyTorch Check the GitHub repo (https://github.com/peace-Van/ClimViz) for details
Files
Steps to reproduce
In short, DeepEcoClimate is derived in four steps (first three steps in MATLAB): - A neural network trained to map climate normal data to land cover - Land cover related climate features obtained from an intermediate layer of the network, PCA-ed and a self-organizing map trained in the reduced feature space - Analyze the SOM structure, manually merge clusters - Migrate to PyTorch via knowledge distillation Details may be revealed in a YouTube video series later.