Spectral Contrast Features (SCF): reference implementation

Reference implementation accompanying:

Spectral Contrast Features: A Bin-Difference Approach for NIR Calibration.

SCF constructs predictive features as differences between the mean intensities of paired spectral bins, whose positions, widths, and sparsity are jointly optimised by a genetic algorithm and passed to PLSR.

What is here

File	Manuscript reference
scf/features.py	Contrast-feature construction: Equation (1); chromosome encoding [s1, w1, s2, w2, active] (Section 2.3)
scf/ga.py	Genetic-algorithm optimisation of the penalised fitness : Equation (2); default hyperparameters reproduce Table 2 (Section 2.4)
scf/plsr.py	PLSR with latent-variable selection by inner cross-validation, "up to 10 latent variables" (Section 2.4); full-spectrum baseline uses the same selection so it is never fixed at an arbitrary component count
scf/evaluation.py	Nested cross-validation and metrics : Section 2.5
scf/preprocessing.py	Wavelength truncation, Savitzky–Golay derivatives, SNV : Section 2.2
examples/wheat_reproduce.py	Deterministic reproduction of the wheat SG² result (Table 3) and cross-instrument transfer (Section 3.1.5) from the GA-selected feature set
examples/wheat_scf_example.py	End-to-end run that rediscovers the contrast features with the genetic algorithm
selftest.py	Runs the whole pipeline on synthetic data : no dataset required

Method in one page

A contrast feature is

f_k = mean( X[:, s1 : s1+w1] )  -  mean( X[:, s2 : s2+w2] )

A candidate solution (chromosome) holds up to P such feature pairs; a binary active gene per pair lets the search choose how many features to use. The GA minimises

fitness = RMSE_cv( X_train, y | active SCF features )  +  lambda * n_active

by tournament selection, feature-level crossover, and per-gene mutation, with elitism. Performance is estimated by a nested cross-validation in which the GA is re-run inside each outer training partition, so the regions selected in a fold never inform that fold's error estimate.

Install and run

pip install -r requirements.txt

# No data needed : verifies the implementation runs:
python selftest.py

# Reproduce the wheat headline result (seconds; uses the GA-selected features):
python examples/wheat_reproduce.py --data-dir /path/to/wheat

# Rediscover the features from scratch with the genetic algorithm:
python examples/wheat_scf_example.py --data-dir /path/to/wheat          # Table 2 settings
python examples/wheat_scf_example.py --data-dir /path/to/wheat --quick  # fast demo

Datasets are public third-party benchmarks and are not redistributed here; see data/README.md for sources and the expected file format.

Reproducibility note

The genetic algorithm is stochastic. The exact bins and metrics vary slightly between runs and random seeds; this is expected and does not affect the method's conclusions. wheat_reproduce.py uses the exact feature set selected on the calibration data and is deterministic : its three features decode to the wavelength regions reported in Section 3.1.2:

F1: [1010.0-1019.5] - [1000.5-1008.0] nm   (N-H second overtone)
F2: [978.5-988.0]   - [898.5-908.0]   nm   (water O-H second overtone)
F3: [855.0-862.0]   - [1012.0-1021.5] nm

The datasets analysed are public third-party benchmarks available from their original sources.

License

MIT : see LICENSE.