Skip to content

preprocessing.md

Latest commit

 

History

History
393 lines (283 loc) · 7.77 KB

preprocessing.md

File metadata and controls

393 lines (283 loc) · 7.77 KB
Raw

Preprocessing and correction

prCARS includes preprocessing, background estimation, correction, denoising, and optional phase-correction utilities for CARS/BCARS spectra.

These steps are important because CARS/BCARS spectra are affected by non-resonant background, noise, baseline effects, and instrument-related distortions before Raman-like information can be recovered.

Typical preprocessing workflow

A typical prCARS workflow is:

Load CARS/BCARS spectrum
Apply optional phase-matching correction
Apply optional denoising
Estimate background
Apply background correction
Run phase retrieval
Apply optional auto-phase correction
Inspect Raman-like output

In code, this is usually handled through the Pipeline interface:

import prcars as ca

pipeline = ca.Pipeline(
    method="kk",
    background="als",
    correction="divide",
    denoise="savgol",
    auto_phase=True,
    silent_region=(2700, 2730),
)

result = pipeline.run(wavenumbers, cars_raw)

Background estimation

Background estimation tries to approximate the smooth non-resonant or baseline-like contribution in the measured CARS/BCARS spectrum.

prCARS currently supports:

Method Description
als Asymmetric least-squares background estimation
polynomial Polynomial background fitting
snip SNIP-style background estimation
rolling_ball Rolling-ball style smooth background estimation
none No background estimation

ALS background

ALS is useful when the background is smooth and the signal contains sharper spectral features.

Example:

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="als",
    correction="divide",
)

Use ALS when:

  • the background is smooth
  • peaks are sharper than the baseline
  • a flexible baseline estimate is needed

Polynomial background

Polynomial fitting is a simple background model.

Example:

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="polynomial",
    correction="subtract",
)

Use polynomial background when:

  • the background is globally smooth
  • a simple interpretable baseline is preferred
  • the spectral window is not too complex

SNIP background

SNIP-style background estimation can be useful for spectra with slowly varying background and peak-like structures.

Example:

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="mem",
    background="snip",
    correction="divide",
)

Use SNIP when:

  • peak-like structures should be preserved
  • the background is broad and smooth
  • you want an alternative to ALS

Rolling-ball background

Rolling-ball background estimation is useful for smooth baseline removal.

Example:

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="rolling_ball",
    correction="divide",
)

Use rolling-ball background when:

  • the background is smooth
  • local baseline behavior matters
  • you want a robust non-parametric background estimate

Background correction modes

After estimating a background, prCARS can correct the measured spectrum before retrieval.

Correction Description
subtract Subtracts the estimated background from the signal
divide Divides the signal by the estimated background
sqrt_divide Applies square-root division, useful for amplitude-like correction
none Does not apply correction

Subtract correction

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="polynomial",
    correction="subtract",
)

Use subtract when the background behaves like an additive baseline.

Divide correction

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="als",
    correction="divide",
)

Use divide when the background behaves more like a multiplicative or normalization factor.

This is often useful in CARS/BCARS workflows because the non-resonant contribution can affect the overall intensity envelope.

Square-root divide correction

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="als",
    correction="sqrt_divide",
)

Use sqrt_divide when working closer to an amplitude-domain correction.

Denoising

prCARS includes optional denoising before retrieval.

Method Description
savgol Savitzky-Golay smoothing
wiener Wiener filtering
wavelet Wavelet-based denoising
none No denoising

Savitzky-Golay denoising

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    denoise="savgol",
)

Use Savitzky-Golay denoising when:

  • noise is moderate
  • peak shapes should be preserved
  • a simple smoothing method is enough

Wiener denoising

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    denoise="wiener",
)

Use Wiener denoising when:

  • the noise is approximately stationary
  • adaptive smoothing may help
  • a simple signal-processing baseline is needed

Wavelet denoising

Wavelet denoising requires the optional wavelet dependency.

Install with:

python -m pip install -e ".[wavelet]"

Then use:

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    denoise="wavelet",
)

Use wavelet denoising when:

  • spectra contain multi-scale noise
  • you want a stronger denoising option
  • preserving local spectral structure is important

Auto-phase correction

prCARS supports optional automatic phase correction using a silent spectral region.

Example:

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    auto_phase=True,
    silent_region=(2700, 2730),
)

Use auto-phase correction when:

  • a silent or low-signal region is known
  • the retrieved Raman-like signal has a phase offset
  • you want to reduce baseline-like phase artifacts

The selected silent region should be chosen carefully based on the spectral window and expected chemistry.

Recommended starting configurations

Simple KK baseline

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="none",
    correction="none",
    denoise="none",
)

Use this first to check raw retrieval behavior.

Practical KK workflow

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="rolling_ball",
    correction="divide",
    denoise="savgol",
)

This is a good practical starting point for synthetic examples.

Stronger preprocessing workflow

result = ca.retrieve(
    wavenumbers,
    cars_raw,
    method="kk",
    background="als",
    correction="divide",
    denoise="wavelet",
    auto_phase=True,
    silent_region=(2700, 2730),
)

Use this when spectra are noisier or require stronger correction.

Notes

Preprocessing choices can strongly affect retrieval quality.

For fair method comparison, always report:

  • retrieval method
  • background method
  • correction mode
  • denoising method
  • phase-correction settings
  • spectral window
  • whether synthetic or real data were used

prCARS is intended for research, education, benchmarking, and portfolio demonstration. It is not intended for clinical diagnosis or deployment.