[Python] Correct scaling of data before KDE with sklearn

I'm using sklearn.neighbors.KernelDensity on geospatial data. I noticed that the bandwidth estimation methods are not considering the spatial extent of the data, only the number of samples and features. The sklearn documentation or the sklearn tutorials don't seem to mention that data should be scaled.

Hence my question: What is the an appropriate way to scale geospatial data before fitting a KDE?

Here is a minimal example:

import numpy as np
from sklearn.neighbors import KernelDensity
import plotly.graph_objects as go
from plotly.subplots import make_subplots

fig = make_subplots(rows=1, cols=2)

X_uniform = np.random.rand(100, 2)
X_upscaled = X_uniform * 100

for i,X in enumerate([X_uniform, X_upscaled]):
kde = KernelDensity(kernel='gaussian', bandwidth="scott").fit(X)
print(f"bw: {kde.bandwidth_}")

# Create grid
x_grid = np.linspace(X[:,0].min(), X[:,0].max(), 50)
y_grid = np.linspace(X[:,1].min(), X[:,1].max(), 50)
X,Y = np.meshgrid(x_grid, y_grid)
xy = np.vstack([X.ravel(), Y.ravel()]).T

z = np.exp(kde.score_samples(xy)).reshape(X.shape)
fig.add_trace(go.Contour(z=z, x=x_grid, y=y_grid), row=1, col=i+1)

fig.show()

Continue reading...