Hydro-Climatic biomes: a multi-task learning approach.

This repository contains the source codes used in the following publication:

• Papagiannopoulou, C., Miralles, D. G., Demuzere, M., Verhoest, N. E. C., and Waegeman, W., 2018. Global hydro-climatic biomes identified via multi-task learning, Geosci. Model Dev., 11, 4139-4153.
• The code itself can be referenced by the following DOI:

Prerequisites

The script is mainly built using the following packages:

python 2.7.13 | Anaconda 5.0.0 (64-bit) (or higher)
scikit-learn

Running the tests

In order to test the framework, a folder with 7 .csv files is provided (test). This folder contains the datasets for the locations {(-29.5,144.5), (-29.5,145.5), (-29.5,147.5), (-30.5,143.5), (29.5, 115.5), (29.5, 112.5), (29.5, 113.5)}. For different locations, the user should modify line 23 of the script main.py, providing the right coordinates. In general, the user can use the datasets provided on the SAT-EX data page in order to run the framework at global scale. In this case, line 23 of the script main.py should be modified into "coords = joblib.load('./coords.pkl')". The file coords.pkl is also provided and it contains the coordinates for all the land pixels (on a 1 x 1° resolution).

To execute the script, use the following commands:

python main.py value_for_parameter_h value_for_parameter_lambda folder_for_csv_files outpath txt_with_variable_names pkl_file_with_the_unused_feature_indices

Where:

• value_for_parameter_h: lower that the initial dimension of the dataset observations (< 3209), a value between 8-12 is recommended for the global dataset, while for the test (small) example of the 7 pixels, the value of 2 or 3 is a good option. This value indicates the number of clusters.

• value_for_parameter_lambda: a value around 10 is recommended

• folder_for_csv_files: e.g., path to the test folder "./test/"

• outpath: path to the output folder

• txt_with_variable_names: given, i.e., "./vars_all.txt"

• pkl_file_with_the_unused_feature_indices: given, i.e., "./rmvdatasets.pkl"

Example:

python main.py 2 10 ./test/  ./out/ ./vars_all.txt ./rmvdatasets.pkl

Note

Files main.py and my_solver.py should be stored in the same folder.

Output

The output of the method is stored in the folder outpath:

The method stores the matrices W, U, V and Theta. For more details see the paper. The clustering algorithm runs on the V matrix. The result matrices can be loaded with the function joblib.load() as numpy arrays.

For a simple visualization of the matrix V, use imshow(). Locations with coherent climate-vegetation interactions should have similar values to the components (columns) of this matrix (V matrix).