Paleo-Climate Precipitation Reconstruction using Bayesian Machine Learning
The main goal of the project is to model global precipitation using geological indicators (including location on continent and sedimentary deposits).
We have available to us the following datasets:
- a database of non-regularly distributed climate-sensitive sedimentary deposits across geologic time (separated on scales of 10-20 million years) and their associated reconstructed geographical location
data/LithData_PaleoXY_Matthews2016_20180226.csv - a dictionary used to translate the lithology codes
data/LithologyCodes.csv - global rainfall modelled at 2.5 deg resolution during Miocene (Herold+ 2012)
data/PRECT_Average_annum.grd - GPlates simulation of continental plates and topology across geologic time (local installation)
- Check that you have necessary packages installed (see
requirements.txt) - download the
query_paleogeography/folder from: https://cloudstor.aarnet.edu.au/plus/s/HRkPptZuiZEm8OV - install pyGPlates (instructions at http://www.gplates.org/docs/pygplates/index.html )
- install GMT
- Note that both the pyGPlates and query_paleogeography folders will need to be in the PYTHONPATH (the pyGPlates documentation talks about this).
Run
python data_preprocess_miocene.py
This should take about 4 minutes. The code uses the deposit database, lithology dictionary, and precipitation simulation output in the data/ folder and creates:
- gridded input data maps in
data/images/ - input data file:
data/learning_data_miocene_YYY.csv
where
- YYY is either
landordeposit(corresponding todata_subsetthe user-set variable within the code)- for files with identifier
land, the dataset will include all points on land (or shallow marine areas that contain deposits) - for files with identifier
deposit, the dataset will include all points where some type of deposit has been found
- for files with identifier
Note: these files will be created locally but are not to be pushed to the online repository
The plate reconstruction queried in this code is designed specifically for the Miocene and is consistent with the (Miocene) precipitation output. Therefore the output csv file is most appropriate for learning.
Run
python data_predictors_anytime.py
This code uses the deposit database and lithology dictionary in the data/ folder and creates:
- gridded input data maps in
data/images/ - input data files:
data/predictor_data_XXX_YYY.csv
where
- XXX is the each geological epoch listed in the deposit database between 6Ma (lower limit for the plate reconstruction query) and 251Ma (Permian)
- YYY is either
landordeposit- for files with identifier
land, the dataset will include all points on land (or shallow marine areas that contain deposits) - for files with identifier
deposit, the dataset will include all points where some type of deposit has been found
- for files with identifier
Note: these files will be created locally but are not to be pushed to the online repository
Step 3: Bayesian model: Predict missing deposit and estimate precip using Gaussian Process - Gibs Sampler based MCMC framework in Matlab
Note: This code is adapted from GPplus. Changes include:
-
references to new variable target_desc instead of Crime
-
tba
-
gridded prediction and uncertainty maps for deposits (coal, evaporites, glacial) and precip using python matplotlib
- R. Chandra, D. Muller, N. Butterworth, and S. Cripps, "Precipitation reconstruction from climate-sensitive lithologies using Bayesian machine learning", Environmental Modelling & Software, 2021 (To Appear).