When spectral libraries are too complex to search: Evolutionary subset selection for domain-adaptive calibration.

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Large spectral libraries offer potential for cost-effective spectroscopy-based sensing.
• Developing accurate predictive models requires domain-specific training samples.
• Identifying relevant samples from large libraries is a significant challenge.

Purpose of the Study:

• Introduce gesearch, a non-linear evolutionary algorithm for selecting target-domain-relevant training samples.
• Enable the development of accurate and interpretable quantitative models using spectral libraries.
• Address the challenge of sample selection for cross-domain spectroscopy applications.

Main Methods:

• Gesearch algorithm for non-linear evolutionary sample selection.
• Application to a North American infrared soil spectral library.
• Cross-domain model development for total carbon prediction in the Democratic Republic of the Congo.

Main Results:

• Gesearch successfully extracted a relevant subset of spectral samples.
• Simple linear models using gesearch samples showed superior accuracy.
• Outperformed established methods like LOCAL, Cubist, CNNs, and PLS regression in a cross-domain test.

Conclusions:

• Gesearch offers a practical framework for utilizing large spectral libraries with limited target data.
• Enables accurate, compact, and interpretable calibration models by selecting spectrally coherent samples.
• The method is effective even when only unlabelled target-domain spectra are available.