Padding data

Several baseline algorithms, notably smoothing and morphological algorithms, pad the input data in order to reduce edge effects from calculations. Padding is performed by pad_edges(), which is a thin wrapper around numpy.pad() that also allows linear extrapolation.

import matplotlib.pyplot as plt
import numpy as np
from scipy.ndimage import uniform_filter1d

from pybaselines.utils import pad_edges


half_window = 80  # for 161 point moving average window
num_points = 1000
x = np.linspace(0, 1000, num_points)
# use an exponentially decaying line to get a large edge effect one
# one side and a small edge effect on the other
line = 10 * np.exp(-x / 150)
noise = np.random.default_rng(0).normal(0, 0.5, num_points)
y = line + noise
pad_len = 2 * half_window + 1

plt.plot(pad_edges(y, pad_len, mode='constant', constant_values=np.nan), label='original')
for i, pad_mode in enumerate(('reflect', 'edge', 'extrapolate'), 1):
    padded_y = pad_edges(y, pad_len, pad_mode, extrapolate_window=int(0.1 * num_points))
    # offset each plot to show differences
    plt.plot(padded_y + 5 * i, label=pad_mode)

plt.legend()

To show the effects that padding can have, the data is smoothed with a moving average. Poor padding produces more severe edge effects from the smoothing.

plt.figure()
plt.plot(y, label='original')
for i, pad_mode in enumerate(('reflect', 'edge', 'extrapolate'), 1):
    padded_y = pad_edges(y, pad_len, pad_mode, extrapolate_window=int(0.1 * num_points))
    plt.plot(
        uniform_filter1d(padded_y, 2 * half_window + 1)[pad_len:num_points + pad_len],
        label=pad_mode
    )

plt.legend()
plt.show()