FTIR Baseline Correction: Methods, Parameters and Common Mistakes

Baseline correction is probably the preprocessing step with the biggest influence on an FTIR analysis. A poorly corrected baseline changes band heights and areas, distorts peak ratios and makes two spectra of the same material look different. This guide explains where sloping or curved baselines come from, which methods exist to correct them, and how to tell whether you have done it right.

Why the baseline drifts

In an ideal spectrum, regions where the sample does not absorb would sit at zero absorbance. In practice, offsets come from:

• Light scattering: in KBr pellets with large or poorly ground particles, or in turbid samples. It produces a baseline that rises towards high wavenumbers, because scattering is stronger at short wavelengths.
• ATR contact: variations in pressure or contact between sample and crystal change the overall intensity and can tilt the spectrum.
• Instrument drift: temperature or detector changes between the background and sample acquisitions.
• Interference fringes in thin films: a regular sinusoidal ripple. Strictly speaking this is not a baseline, and the usual methods do not remove it well.

Atmospheric CO₂ (≈ 2350 cm⁻¹) and water vapour bands are not a baseline problem either: they are solved with a good background or by purging the instrument.

First: work in absorbance

Always correct the baseline in absorbance, not transmittance. The Beer–Lambert law is additive in absorbance, so an offset baseline can be subtracted; in transmittance the effect is multiplicative and subtraction has no physical meaning.

A = −log₁₀(T)        (T as a fraction)
A = 2 − log₁₀(%T)    (T as a percentage)

Correction methods

1. Anchor points (linear or polynomial)

You pick several points where you know the sample does not absorb, fit a piecewise straight line or a polynomial through them, and subtract it.

• Pros: full control, easy to justify in a paper.
• Cons: depends on choosing the points well; poorly reproducible between users. High-order polynomials (> 3) oscillate and can create or erase bands.

2. Rubberband (convex hull)

Picture a rubber band stretched underneath the spectrum: it touches the minima and is subtracted.

• Pros: automatic, practically parameter-free, fast.
• Cons: works well with monotonic or convex baselines, but if the true baseline curves upwards it can clip broad bands such as O–H (3200–3600 cm⁻¹).

3. Asymmetric least squares (AsLS)

Proposed by Eilers and Boelens (2005). It fits a smooth curve that penalises points lying below it much more than those above it, so it hugs the base and “ignores” the peaks. It has two parameters:

Parameter	What it controls	Typical range
λ (smoothness)	How much the baseline is allowed to bend	10² – 10⁹
p (asymmetry)	Weight of points above the baseline	0.001 – 0.1

• Pros: very versatile with curved baselines; reproducible if you fix the parameters.
• Cons: too low a λ makes the baseline follow broad bands and shrink them; too high a λ leaves curvature uncorrected.

4. Derivatives (Savitzky–Golay)

The first derivative removes a constant offset and the second derivative also removes a linear slope. They also help resolve overlapping bands.

• Pros: no baseline to choose; useful in chemometrics and for locating overlapping peaks.
• Cons: amplifies noise (needs smoothing) and changes band shapes, so it is not suitable for measuring heights or areas directly.

Which one to choose?

Situation	Recommended method
Simple spectra, few samples, publication	Linear anchor points
Many spectra with an almost straight base	Rubberband
Curved baselines (scattering, KBr, biological samples)	AsLS
Classification, PCA or heavily overlapping peaks	Savitzky–Golay derivatives

Recommended preprocessing order

1. Convert to absorbance.
2. ATR correction, if you will compare with transmission spectra or libraries measured in transmission.
3. Baseline correction.
4. Mild smoothing (for example Savitzky–Golay with a short window), only if the noise requires it.
5. Normalisation: to a reference band that does not change between samples, by area or by vector, depending on the goal.

Apply the same parameters to every spectrum you are going to compare. Changing method or λ between samples introduces differences that do not come from the material.

How to check the correction is good

• Overlay the computed baseline on the original spectrum before subtracting: it should run through the absorption-free regions, not cut through bands.
• Look at band-free regions: after correction they should be close to zero. Be careful with 2300–1900 cm⁻¹ on a diamond ATR, where the crystal itself absorbs.
• No strongly negative absorbance: it usually means the baseline runs above the spectrum.
• Check broad bands (O–H, N–H, water in hydrogels): they are the first casualties of over-aggressive correction.

Practical example: comparing the carbonyl band of PLA and PCL

The carbonyl group (C=O) appears around 1750 cm⁻¹ in PLA and 1720 cm⁻¹ in PCL. If you measure that band’s area to follow, for instance, the composition of a PLA/PCL blend or the degradation of a film:

• With an uncorrected sloping baseline, the band area includes a “background” trapezoid that changes from sample to sample depending on ATR contact.
• With a linear two-point correction on either side of the band (for example ~1850 and ~1650 cm⁻¹, adjusted to your spectrum), the area reflects only the carbonyl absorption and comparisons between samples make sense.
• If you also normalise to a band that does not change during the process, you remove the effect of thickness or contact differences.

Common mistakes

1. Correcting in transmittance instead of absorbance.
2. Using different parameters for each spectrum in the same comparison.
3. High-order polynomials that create artificial ripples.
4. Measuring areas on derivative spectra.
5. Over-smoothing before correction, broadening bands and merging nearby peaks.
6. Mistaking interference fringes for baseline: they need a different treatment.

MIND4SCE-Infrared lets you import your spectra (.csv, .xlsx, .dpt, .txt, .spc), apply baseline correction, Savitzky–Golay smoothing and normalisation, overlay samples and compare them against a library of 90+ reference compounds — all in the browser.