Machine Learning for Species Distribution analysis

Python code for implementing and testing various models on an analysis of the distribution of species around the world

The following code is separated into three different folders; src, results and data.

Here is a brief overview of the structure of the code layout:

• /src: contains all code used to generate results from the project, separated into two different folders for implementing machine learning models and analysis/data preparation.

  • models: models contains all the models and implementations investigated
  • data preparation: code used to investigate data; such as finding sparse/dense populations, investigating top species and analyzing bioclimatic and climate data appropriately for training

• /data: contains all data, given and generated, used to train and test models.

• /results: found results and visualization methods (via seaborn and matplotlib)

Requirements

To run the code below a python virtual environment with the following packages installed will need to be setup:

• matplotlib
• geopandas
• numpy
• shapely
• sklearn
• seaborn
• PyTorch
• geopy
• pandas

Running code and obtaining results

/src :

• /models:

  To run all scripts for the 2 feature trained models:

      models_2_feature.sh

  To run all scripts for the 8 feature trained models:

      models_8_feature.sh

  A description of running the code is included below:

  • Feed-forward neural network algorithms:

  To run the results for the Feed-Forward neural network trained on two features (latitude and longitude) run 'neural_network.py' in the terminal as: python neural_network.py a variable 'p' is included at the top of the script; change between p = "analyze" or p = "plot" to either run an analysis of the neural network model and print AUCROC, AUCPR, F2 and cohen kappa measures or to plot the predicted distribution of a species respectively.

  To run the results for the Feed-Forward neural network trained on eight features run 'neural_network_5.py' in the terminal as: python neural_network_5.py similary to the 2 feature model a variable 'p' is included at the top of the script; change between p = "analyze" or p = "plot" to either run an analysis of the neural network model and print AUCROC, AUCPR, F2 and cohen kappa measures or to plot the predicted distribution of a species respectively.

  • Random Forest algorithms:

  To run the results for the Random Forest trained on 2 features, run random_forest.py. This will return all the results and evaluation for the random forest, as well as a plot for a given species.

  To run the results for the model trained on 8 features, run v2_random_forest.py. This will return the evaluation. Note that only F2 was added to the report.

  • Logistic regression algorithms:

  To obtain the results for logistic regression trained on 2 features run log_regression_model.py. For the 8 feature model, run lr_8_features.py. To obtain the distribution plot run lr_plot.py.

  • K-Nearest Neighbour algorithms: To find the optimal value of k run either knn_optim.py (2 feature) or (knn_8_features_optim.py). To obtain the results for knn trained on 2 features run knn_model.py. For the 8 feature model, run knn_8_features.py. To obtain the distribution plot run knn_plot.py.

  • Gaussian algorithm:

  The Gaussian model was not used in the report, but the code can still be used to do some basic predictions on the 2 feature data. Running gaussian_model.py will generate the predicted top 3 species present at the coordinate given by Edinburgh city center. To obtain a distribution plot like for other models run gaussian_contours.py.

• /data preparation:

  To run all scripts for the data preparation tasks:

      data_prep.sh

  Note, all data has already been generated and saved into the /data folder, however instructions to run the files to generate the data are given below:

  • different species distribution types:

  To obtain the results on the different species distribution types (dense vs sparse, and largest vs smallest span respectively) run 'species_distribution_analysis.py' in the terminal as: python species_distribution_analysis.py -d density python species_distribution_analysis.py -d distribution The data is then saved into a .csv file

  • generate top species continent analysis:

  To obtain the species' most common continent run create_df_continent.py. Dataframe is then saved into a .csv file.

  The top_species_analysis.py file does a similar procedure but only creates a file with the species with 2000 data points in the training data. What we denominated as the "top species".

  • generate 8 feature data:

  To generate the 8 feature data by extracting from the data available from WorldClim run 'species_distribution_analysis.py' as: python species_distribution_analysis.py -d extra The data is then saved into a .csv file

  • generate temperature anomaly scores: Running temperature_anomaly.py will generate the plot of the temperature anomaly seen in the appendix, along with the scores (score.npy) and associated locations for those scores (scores_coords.npy)