物种分布模型

物种地理分布的模拟是保护生物学中的一个重要问题。在这个例子中，我们模拟了两种南美哺乳动物的地理分布，给出了过去的观测结果和14个环境变量。由于我们只有正的例子(没有不成功的观测结果)，所以我们把这个问题归结为密度估计问题，并使用 sklearn.svm.OneClassSVM作为我们的建模工具。数据集由Phillips et. al. (2006)提供。如果有，该示例使用basemap绘制南美洲的海岸线和国家边界。

这两个物种是：

参靠

________________________________________________________________________________
Modeling distribution of species 'bradypus variegatus'
 - fit OneClassSVM ... done.
 - plot coastlines from coverage
 - predict species distribution

 Area under the ROC curve : 0.868443
________________________________________________________________________________
Modeling distribution of species 'microryzomys minutus'
 - fit OneClassSVM ... done.
 - plot coastlines from coverage
 - predict species distribution

 Area under the ROC curve : 0.993919

time elapsed: 29.37s
# Authors: Peter Prettenhofer <peter.prettenhofer@gmail.com>
#          Jake Vanderplas <vanderplas@astro.washington.edu>
#
# License: BSD 3 clause

from time import time

import numpy as np
import matplotlib.pyplot as plt

from sklearn.utils import Bunch
from sklearn.datasets import fetch_species_distributions
from sklearn import svm, metrics

# if basemap is available, we'll use it.
# otherwise, we'll improvise later...
try:
    from mpl_toolkits.basemap import Basemap
    basemap = True
except ImportError:
    basemap = False

print(__doc__)


def construct_grids(batch):
    """Construct the map grid from the batch object

    Parameters
    ----------
    batch : Batch object
        The object returned by :func:`fetch_species_distributions`

    Returns
    -------
    (xgrid, ygrid) : 1-D arrays
        The grid corresponding to the values in batch.coverages
    """
    # x,y coordinates for corner cells
    xmin = batch.x_left_lower_corner + batch.grid_size
    xmax = xmin + (batch.Nx * batch.grid_size)
    ymin = batch.y_left_lower_corner + batch.grid_size
    ymax = ymin + (batch.Ny * batch.grid_size)

    # x coordinates of the grid cells
    xgrid = np.arange(xmin, xmax, batch.grid_size)
    # y coordinates of the grid cells
    ygrid = np.arange(ymin, ymax, batch.grid_size)

    return (xgrid, ygrid)


def create_species_bunch(species_name, train, test, coverages, xgrid, ygrid):
    """Create a bunch with information about a particular organism

    This will use the test/train record arrays to extract the
    data specific to the given species name.
    """
    bunch = Bunch(name=' '.join(species_name.split("_")[:2]))
    species_name = species_name.encode('ascii')
    points = dict(test=test, train=train)

    for label, pts in points.items():
        # choose points associated with the desired species
        pts = pts[pts['species'] == species_name]
        bunch['pts_%s' % label] = pts

        # determine coverage values for each of the training & testing points
        ix = np.searchsorted(xgrid, pts['dd long'])
        iy = np.searchsorted(ygrid, pts['dd lat'])
        bunch['cov_%s' % label] = coverages[:, -iy, ix].T

    return bunch


def plot_species_distribution(species=("bradypus_variegatus_0",
                                       "microryzomys_minutus_0")):
    """
    Plot the species distribution.
    """
    if len(species) > 2:
        print("Note: when more than two species are provided,"
              " only the first two will be used")

    t0 = time()

    # Load the compressed data
    data = fetch_species_distributions()

    # Set up the data grid
    xgrid, ygrid = construct_grids(data)

    # The grid in x,y coordinates
    X, Y = np.meshgrid(xgrid, ygrid[::-1])

    # create a bunch for each species
    BV_bunch = create_species_bunch(species[0],
                                    data.train, data.test,
                                    data.coverages, xgrid, ygrid)
    MM_bunch = create_species_bunch(species[1],
                                    data.train, data.test,
                                    data.coverages, xgrid, ygrid)

    # background points (grid coordinates) for evaluation
    np.random.seed(13)
    background_points = np.c_[np.random.randint(low=0, high=data.Ny,
                                                size=10000),
                              np.random.randint(low=0, high=data.Nx,
                                                size=10000)].T

    # We'll make use of the fact that coverages[6] has measurements at all
    # land points.  This will help us decide between land and water.
    land_reference = data.coverages[6]

    # Fit, predict, and plot for each species.
    for i, species in enumerate([BV_bunch, MM_bunch]):
        print("_" * 80)
        print("Modeling distribution of species '%s'" % species.name)

        # Standardize features
        mean = species.cov_train.mean(axis=0)
        std = species.cov_train.std(axis=0)
        train_cover_std = (species.cov_train - mean) / std

        # Fit OneClassSVM
        print(" - fit OneClassSVM ... ", end='')
        clf = svm.OneClassSVM(nu=0.1, kernel="rbf", gamma=0.5)
        clf.fit(train_cover_std)
        print("done.")

        # Plot map of South America
        plt.subplot(1, 2, i + 1)
        if basemap:
            print(" - plot coastlines using basemap")
            m = Basemap(projection='cyl', llcrnrlat=Y.min(),
                        urcrnrlat=Y.max(), llcrnrlon=X.min(),
                        urcrnrlon=X.max(), resolution='c')
            m.drawcoastlines()
            m.drawcountries()
        else:
            print(" - plot coastlines from coverage")
            plt.contour(X, Y, land_reference,
                        levels=[-9998], colors="k",
                        linestyles="solid")
            plt.xticks([])
            plt.yticks([])

        print(" - predict species distribution")

        # Predict species distribution using the training data
        Z = np.ones((data.Ny, data.Nx), dtype=np.float64)

        # We'll predict only for the land points.
        idx = np.where(land_reference > -9999)
        coverages_land = data.coverages[:, idx[0], idx[1]].T

        pred = clf.decision_function((coverages_land - mean) / std)
        Z *= pred.min()
        Z[idx[0], idx[1]] = pred

        levels = np.linspace(Z.min(), Z.max(), 25)
        Z[land_reference == -9999] = -9999

        # plot contours of the prediction
        plt.contourf(X, Y, Z, levels=levels, cmap=plt.cm.Reds)
        plt.colorbar(format='%.2f')

        # scatter training/testing points
        plt.scatter(species.pts_train['dd long'], species.pts_train['dd lat'],
                    s=2 ** 2, c='black',
                    marker='^', label='train')
        plt.scatter(species.pts_test['dd long'], species.pts_test['dd lat'],
                    s=2 ** 2, c='black',
                    marker='x', label='test')
        plt.legend()
        plt.title(species.name)
        plt.axis('equal')

        # Compute AUC with regards to background points
        pred_background = Z[background_points[0], background_points[1]]
        pred_test = clf.decision_function((species.cov_test - mean) / std)
        scores = np.r_[pred_test, pred_background]
        y = np.r_[np.ones(pred_test.shape), np.zeros(pred_background.shape)]
        fpr, tpr, thresholds = metrics.roc_curve(y, scores)
        roc_auc = metrics.auc(fpr, tpr)
        plt.text(-35, -70, "AUC: %.3f" % roc_auc, ha="right")
        print("\n Area under the ROC curve : %f" % roc_auc)

    print("\ntime elapsed: %.2fs" % (time() - t0))


plot_species_distribution()
plt.show()

脚本的总运行时间：(0分29.592秒)

Download Python source code:plot_species_distribution_modeling.py

Download Jupyter notebook:plot_species_distribution_modeling.ipynb