SoilSpecTfm

SoilSpecTfm corrects, smooths and resamples soil infrared spectra before modeling. It covers scatter correction, Savitzky-Golay smoothing and derivatives, wavelet denoising, conversion to absorbance, resampling and trimming. Each transformer follows the scikit-learn API. You can chain them in a Pipeline, tune their parameters with GridSearchCV, and put them in front of any scikit-learn model.

SoilSpecData loads the Open Soil Spectral Library (OSSL) as spectra and soil properties ready for these transformers. The SoilSpecTfm documentation covers the full API.

Installation

pip install -U soilspectfm

To install the development version, with nbdev and SoilSpecData, run this in the project root with uv:

uv sync --extra dev

Features

soilspectfm.datasets loads the two toy datasets, and soilspectfm.visualization plots spectra.

Quick start

from soilspectfm.all import * imports every transformer, the toy dataset loaders and the plotting functions. Load the toy MIR spectra that ship with the package. Spectrum edges are often noisy, and one toy spectrum has a faulty first point, so Trim keeps the wavenumbers from 650 cm⁻¹. Then preprocess the spectra with standard normal variate (SNV), wavelet denoising and a first derivative:

from soilspectfm.all import *
from sklearn.pipeline import Pipeline

X, ws = load_toy_mir()
trim = Trim(ws, w_min=650)
X, ws = trim.fit_transform(X), trim.get_wavenumbers()
pipe = Pipeline([
    ('snv', SNV()),
    ('denoise', WaveletDenoise()),
    ('deriv', SavitzkyGolay(window_length=11, polyorder=2, deriv=1))])
X_tfm = pipe.fit_transform(X)
X_tfm.shape

(50, 1676)

Compare the raw and preprocessed spectra:

plot_spectra_comparison(
    X, X_tfm, ws,
    transformed_title='SNV, wavelet denoising and first derivative');

Modeling OSSL data

SoilSpecData gives the spectra X and soil properties y of the OSSL samples. This example predicts cation exchange capacity (CEC) from MIR spectra with partial least squares regression, and scores it with 5-fold cross-validation. The first call to load_ossl downloads about 1 GB:

from soilspecdata.all import *
from sklearn.cross_decomposition import PLSRegression
from sklearn.model_selection import cross_val_score

X, y, ids = load_ossl().mir().xy('cec_usda.a723_cmolc.kg')
model = Pipeline([
    ('snv', SNV()),
    ('deriv', SavitzkyGolay(window_length=11, polyorder=2, deriv=1)),
    ('pls', PLSRegression(n_components=10))])
cross_val_score(model, X, y, cv=5, scoring='r2')

Upgrading from 0.0.x

Version 0.1.0 changes the API of soilspectfm. To update code written for version 0.0.x:

• soilspectfm.core is now soilspectfm.preprocessing. To import every transformer, loader and plotting function at once, use from soilspectfm.all import *.
• SavitzkyGolay replaces TakeDerivative and SavGolSmooth. Its defaults are window_length=15, polyorder=3 and deriv=0, the old SavGolSmooth defaults. Replace SavGolSmooth(...) with SavitzkyGolay(...). Replace TakeDerivative() with SavitzkyGolay(window_length=11, polyorder=1, deriv=1), because the defaults differ.
• MSC no longer accepts n_jobs.
• Resample.fit raises a ValueError when target_x goes outside the range of source_x. Resample.transform now requires fit first.
• soilspectfm.utils is now soilspectfm.datasets. It no longer exports toy_mir_url and toy_noisy_mir_url. The toy datasets ship with the package, and load_toy_mir and load_toy_noisy_mir read them without a network connection.

Contributing

Contributions are welcome. SoilSpecTfm is built with nbdev: the notebooks in nbs/ are the source of the code, the documentation and the tests. After changing a notebook, export the code, run the tests, clean the notebooks and rebuild the README with:

uv run nbdev-prepare

Report bugs and ask questions in GitHub issues.

License

SoilSpecTfm is under the Apache 2.0 licence.