Baseline Correction

Baseline correction removes instrumental artifacts and background signals from FTIR spectra.

Overview

The xpectrass baseline module wraps pybaselines to provide 50+ baseline correction algorithms through a unified interface.

Using FTIRdataprocessing Class (Recommended)

The easiest way to apply baseline correction is through the FTIRdataprocessing class:

from xpectrass import FTIRdataprocessing
from xpectrass.data import load_jung_2018

# Initialize with sample data
df = load_jung_2018().head(80)
ftir = FTIRdataprocessing(df=df, label_column="type")

# Prepare denoised input first (baseline evaluation expects processed spectra)
df_denoised = ftir._get_denoised_data(denoising_method="wavelet", plot=False)

# Step 1: Evaluate baseline methods
rfzn, nar, snr = ftir.find_baseline_method(
    data=df_denoised,
    flat_windows=[(1800, 1900), (2400, 2700)],
    baseline_methods="FTIR",
    n_samples=20,
    plot=False,
)

# Step 2: View evaluation metrics
print(ftir.rfzn_tbl)  # Residual Flatness in Zero Noise
print(ftir.nar_tbl)   # Negative Absorbance Ratio
print(ftir.snr_tbl)   # Signal-to-Noise Ratio

# Step 3: Plot evaluation results
ftir.plot_rfzn_nar_snr(metric_name="RFZN")

# Step 4: Apply the best method
ftir.correct_baseline(data=df_denoised, method="asls", lam=1e6, plot=False)

# Step 5: Access corrected data
corrected_df = ftir.df_corr

Using Utility Functions Directly

For standalone use or custom pipelines:

from xpectrass.utils import baseline_correction, baseline_method_names

# See all available methods
print(baseline_method_names())

# Apply baseline correction to a single spectrum
corrected = baseline_correction(intensities, method='airpls', lam=1e6)

Available Methods

Whittaker-Based Methods

Method	Description	Best For
`asls`	Asymmetric Least Squares	General purpose
`airpls`	Adaptive Iteratively Reweighted Penalized LS	Default choice for FTIR
`arpls`	Asymmetrically Reweighted Penalized LS	Strong baselines
`iasls`	Improved AsLS	Better peak preservation
`psalsa`	Peak-Screening AsLS	Sharp peaks
`aspls`	Adaptive Smoothness Penalized LS	Variable smoothness

Polynomial-Based Methods

Method	Description
`poly`	Standard polynomial fit
`modpoly`	Modified polynomial
`imodpoly`	Iterative modified polynomial
`penalized_poly`	Penalized polynomial
`loess`	Local regression

Morphological Methods

Method	Description
`mor`	Morphological opening
`imor`	Iterative morphological
`mormol`	Morphological and mollification
`rolling_ball`	Rolling ball algorithm
`tophat`	Top-hat transform

Spline-Based Methods

Method	Description
`mixture_model`	Mixture model approach
`irsqr`	Iteratively reweighted spline quantile regression
`pspline_asls`	Penalized spline AsLS

Custom Methods

Method	Description
`median_filter`	Median filter baseline
`adaptive_window`	Adaptive minimum filter

Function Reference

baseline_correction

corrected = baseline_correction(
    intensities,           # 1-D array of intensities
    method='airpls',       # Algorithm name
    window_size=101,       # For custom windowed filters
    poly_order=4,          # For polynomial methods
    clip_negative=True,    # Set negative values to 0
    return_baseline=False, # Return (corrected, baseline) tuple
    **kwargs               # Method-specific parameters
)

Common Parameters by Method

For Whittaker methods (asls, airpls, arpls, etc.):

lam: Smoothness parameter (typically 1e4 to 1e8). Higher = smoother baseline.
p: Asymmetry parameter (typically 0.001 to 0.1). Lower = less peak influence.

For polynomial methods:

poly_order: Polynomial degree (typically 2-6)

Evaluation

Compare baseline methods using RFZN and NAR metrics:

from xpectrass.utils import evaluate_baseline_correction_methods

# Define flat regions (known baseline-only areas)
flat_windows = [(2500, 2600), (3350, 3450)]

# Evaluate all methods
rfzn, nar, snr = evaluate_baseline_correction_methods(
    data=df,
    flat_windows=flat_windows,
    label_column="type",
    exclude_columns=["study", "sample_id", "environmental", "resolution"],
    baseline_methods=["asls", "airpls", "arpls"],
    n_samples=20,
    sample_selection="random",
)

# Lower RFZN and NAR = better baseline correction
print("Best methods by RFZN:", rfzn.mean().sort_values().head())

Metrics

Metric	Full Name	Interpretation
RFZN	Residual Flat-Zone Noise	RMS of corrected signal in known baseline regions. Lower = better.
NAR	Negative Area Ratio	Fraction of negative area. Lower = better.
SNR	Signal-to-Noise Ratio	Peak height / noise. Higher = better.

Visualization

from xpectrass.utils import plot_baseline_correction_metric_boxes

# Visualize RFZN distribution for evaluated methods
plot_baseline_correction_metric_boxes(
    df=rfzn,
    metric_name="RFZN",
)

Recommendations for Plastics

Plastic Type	Recommended Method	Notes
HDPE, LDPE	`airpls`	Strong CH peaks, smooth baseline
PET	`asls` or `airpls`	Complex spectrum
PP	`airpls`	Similar to PE
PS	`airpls`	Aromatic features
PVC	`airpls` or `arpls`	May have strong baseline drift

Example

import numpy as np
from xpectrass.utils import baseline_correction

# Load spectrum
wavenumbers = np.linspace(400, 4000, 3751)
intensities = load_spectrum('HDPE1.csv')

# Apply baseline correction
corrected = baseline_correction(
    intensities,
    method='airpls',
    lam=1e6
)

# Get baseline for visualization
corrected, baseline = baseline_correction(
    intensities,
    method='airpls',
    lam=1e6,
    return_baseline=True
)