Peak Detection in Python for Spectroscopy: find_peaks, Prominence, and Baseline Correction

Spectroscopy plots are rarely clean. Baselines drift, noise hides weaker peaks, and a naive threshold can miss the features that matter. A better workflow is to smooth first, tune peak prominence carefully, and then inspect the detected peaks before you trust the result.

Example workflow

The code below keeps the logic intentionally simple. In real Raman, FTIR, UV-Vis, or chromatogram data, you would usually add baseline correction and maybe fit the strongest peaks with Gaussian or Lorentzian models.

import numpy as np
import matplotlib.pyplot as plt
from scipy.signal import find_peaks, savgol_filter

x = np.linspace(400, 1800, 1200)
signal = np.exp(-0.5 * ((x - 730) / 18) ** 2) + 0.6 * np.exp(-0.5 * ((x - 1100) / 25) ** 2)
signal += 0.03 * np.sin(x / 25) + np.random.default_rng(4).normal(0, 0.03, len(x))

smoothed = savgol_filter(signal, 41, 3)
peaks, props = find_peaks(smoothed, prominence=0.08, width=10)

fig, ax = plt.subplots(figsize=(8, 4.5))
ax.plot(x, signal, color="gray", alpha=0.5, label="Raw")
ax.plot(x, smoothed, color="black", lw=1.5, label="Smoothed")
ax.plot(x[peaks], smoothed[peaks], "o", color="crimson")
ax.legend(frameon=False)
plt.tight_layout()

What matters most

• Prominence is usually more useful than height when the baseline shifts.
• Width helps separate broad real features from narrow noise spikes.
• Peak distances prevent the same feature from being counted twice.
• Always annotate units on the x-axis, especially for spectra.