Rekyt / IUCNN

Train neural networks based on geographic species occurrences, environmental data and existing IUCN Red List assessments to predict the conservation status of "Not Evaluated" or "Data Deficient" species, for any taxon or geographic region of interest.

IUCNN

Batch estimation of species' IUCN Red List threat status using neural networks.

Installation

1. Install IUCNN directly from Github using devtools.

install.packages("devtools")
library(devtools)

install_github("azizka/IUCNN")

2. Since some of IUCNNs functions are run in Python, IUCNN needs to set up a Python environment. This is easily done from within R, using the install_miniconda() function of the package reticulate (this will need c. 3 GB disk space). If problems occur at this step, check the excellent documentation of reticulate.

install.packages("reticulate")
library(reticulate)
install_miniconda()

3. Install the tensorflow python library. If you are using MacOS or Linux it is recommended to install tensorflow using conda:

reticulate::conda_install("r-reticulate","tensorflow=2.4")

If you are using Windows, you can install tensorflow using pip:

reticulate::py_install("tensorflow~=2.4.0rc4", pip = TRUE)

4. Finally install the npBNN python library from Github:

reticulate::py_install("https://github.com/dsilvestro/npBNN/archive/v0.1.10.tar.gz", pip = TRUE)

Usage

There are multiple models and features available in IUCNN. A vignette with a detailed tutorial on how to use those is available as part of the package: vignette("Approximate_IUCN_Red_List_assessments_with_IUCNN"). Running IUCNN will write files to your working directory.

A simple run:

library(tidyverse)
library(IUCNN)

#load example data 
data("training_occ") #geographic occurrences of species with IUCN assessment
data("training_labels")# the corresponding IUCN assessments
data("prediction_occ") #occurrences from Not Evaluated species to prdict

# 1. Feature and label preparation
features <- prep_features(training_occ) # Training features
labels_train <- prep_labels(training_labels) # Training labels
features_predict <- prep_features(prediction_occ) # Prediction features

# 2. Model training
m1 <- train_iucnn(x = features, lab = labels_train)

summary(m1)
plot(m1)

# 3. Prediction
predict_iucnn(x = features_predict,
              model = m1)

With model testing

library(tidyverse)
library(IUCNN)

#load example data 
data("training_occ") #geographic occurrences of species with IUCN assessment
data("training_labels")# the corresponding IUCN assessments
data("prediction_occ") #occurrences from Not Evaluated species to prdict

# Feature and label preparation
features <- prep_features(training_occ) # Training features
labels_train <- prep_labels(training_labels) # Training labels
features_predict <- prep_features(prediction_occ) # Prediction features


# Model testing
# For illustration models differing in dropout rate and number of layers
mod_test <- modeltest_iucnn(features,
                            labels_train,
                            logfile = "model_testing_results-2.txt",
                            model_outpath = "iucnn_modeltest-2",
                            mode = "nn-class",
                            dropout_rate = c(0.0, 0.1, 0.3),
                            n_layers = c("30", "40_20", "50_30_10"),
                            cv_fold = 5,
                            init_logfile = TRUE)

# Select best model
m_best <- bestmodel_iucnn(mod_test,
                          criterion = "val_acc",
                          require_dropout = TRUE)

# Inspect model structure and performance
summary(m_best)
plot(m_best)

# Train the best model on all training data for prediction
m_prod <- train_iucnn(features,
                      labels_train,
                      production_model = m_best,
                      overwrite = TRUE)

# Predict RL categories for target species
pred <- predict_iucnn(features_predict,
                      m_prod)
plot(pred)

Citation

library(IUCNN)
citation("IUCNN")

Zizka A, Silvestro D, Vitt P, Knight T (2020). “Automated conservation assessment of the orchid family with deep learning.” Conservation Biology, 0, 0-0. doi: doi.org/10.1111/cobi.13616 (URL: https://doi.org/doi.org/10.1111/cobi.13616), <URL: https://github.com/azizka/IUCNN>.