A simple Maxent model¶

Here's an end-to-end example for fitting and applying a Maxent model.

This is to demonstrate the simplest pattern of model training. Full model training and evaluation should include creating train/test splits, cross-validation, feature selection, which aren't covered here.

The elapid sample data are occurrence records for Ariolimax buttoni, a species of banana slug native to coastal California, annotated with climate and vegetation data.

In [1]:

# packages
import elapid as ela

import rasterio as rio
import geopandas as gpd
import matplotlib as mpl
import matplotlib.pyplot as plt
from sklearn import metrics
import warnings

# paths
https = "https://earth-chris.github.io/images/research"
vector = f"{https}/ariolimax-buttoni.gpkg"
raster_names = [
    "ca-cloudcover-mean.tif",
    "ca-cloudcover-stdv.tif",
    "ca-leafareaindex-mean.tif",
    "ca-leafareaindex-stdv.tif",
    "ca-surfacetemp-mean.tif",
    "ca-surfacetemp-stdv.tif",
]
rasters = [f"{https}/{raster}" for raster in raster_names]
labels = [raster[3:-4] for raster in raster_names]

# preferences
%matplotlib inline
mpl.style.use('ggplot')
warnings.filterwarnings("ignore")

print(f"Notebook last run with elapid version {ela.__version__}")

# packages import elapid as ela import rasterio as rio import geopandas as gpd import matplotlib as mpl import matplotlib.pyplot as plt from sklearn import metrics import warnings # paths https = "https://earth-chris.github.io/images/research" vector = f"{https}/ariolimax-buttoni.gpkg" raster_names = [ "ca-cloudcover-mean.tif", "ca-cloudcover-stdv.tif", "ca-leafareaindex-mean.tif", "ca-leafareaindex-stdv.tif", "ca-surfacetemp-mean.tif", "ca-surfacetemp-stdv.tif", ] rasters = [f"{https}/{raster}" for raster in raster_names] labels = [raster[3:-4] for raster in raster_names] # preferences %matplotlib inline mpl.style.use('ggplot') warnings.filterwarnings("ignore") print(f"Notebook last run with elapid version {ela.__version__}")

Notebook last run with elapid version 1.0.1

In [2]:

# read the presence data, draw background point samples
presence = gpd.read_file(vector)
background = ela.sample_raster(rasters[0], count=10_000)

# merge datasets and read the covariates at each point location
merged = ela.stack_geodataframes(presence, background, add_class_label=True)
annotated = ela.annotate(merged, rasters, drop_na=True, quiet=True)

# split the x/y data
x = annotated.drop(columns=['class', 'geometry'])
y = annotated['class']

# train the model
model = ela.MaxentModel(transform='cloglog', beta_multiplier=2.0)
model.fit(x, y)

# evaluate training performance
ypred = model.predict(x)
auc = metrics.roc_auc_score(y, ypred)
print(f"Training AUC score: {auc:0.3f}")

# save the fitted model to disk
ela.save_object(model, 'demo-maxent-model.ela')

# read the presence data, draw background point samples presence = gpd.read_file(vector) background = ela.sample_raster(rasters[0], count=10_000) # merge datasets and read the covariates at each point location merged = ela.stack_geodataframes(presence, background, add_class_label=True) annotated = ela.annotate(merged, rasters, drop_na=True, quiet=True) # split the x/y data x = annotated.drop(columns=['class', 'geometry']) y = annotated['class'] # train the model model = ela.MaxentModel(transform='cloglog', beta_multiplier=2.0) model.fit(x, y) # evaluate training performance ypred = model.predict(x) auc = metrics.roc_auc_score(y, ypred) print(f"Training AUC score: {auc:0.3f}") # save the fitted model to disk ela.save_object(model, 'demo-maxent-model.ela')

Training AUC score: 0.981

In [3]:

# partial dependence plots show how model predictions vary across the range of covariates
fig, ax = model.partial_dependence_plot(x, labels=labels, dpi=100)

# partial dependence plots show how model predictions vary across the range of covariates fig, ax = model.partial_dependence_plot(x, labels=labels, dpi=100)

In [4]:

# write the model predictions to disk
output_raster = 'demo-maxent-predictions.tif'
ela.apply_model_to_rasters(model, rasters, output_raster, quiet=True)

# and read into memory
with rio.open(output_raster, 'r') as src:
    pred = src.read(1, masked=True)

# write the model predictions to disk output_raster = 'demo-maxent-predictions.tif' ela.apply_model_to_rasters(model, rasters, output_raster, quiet=True) # and read into memory with rio.open(output_raster, 'r') as src: pred = src.read(1, masked=True)

In [5]:

# plot the suitability predictions
fig, ax = plt.subplots(1, 1, figsize=(7, 6), dpi=100)
plot = ax.imshow(pred, vmin=0, vmax=1, cmap='GnBu')
ax.set_title('$Ariolimax\ buttoni$ suitability')
ax.set_xticks([])
ax.set_yticks([])
cbar = plt.colorbar(plot, ax=ax, label="relative occurrence probability", pad=0.04)
plt.tight_layout()

# plot the suitability predictions fig, ax = plt.subplots(1, 1, figsize=(7, 6), dpi=100) plot = ax.imshow(pred, vmin=0, vmax=1, cmap='GnBu') ax.set_title('$Ariolimax\ buttoni$ suitability') ax.set_xticks([]) ax.set_yticks([]) cbar = plt.colorbar(plot, ax=ax, label="relative occurrence probability", pad=0.04) plt.tight_layout()