Rethinking local spectral modelling: From per-query refitting to model libraries

BACKGROUND: This work introduces liblex, a simple yet powerful algorithmic framework to streamline the use of complex spectral libraries. It converts diffuse reflectance spectroscopy (DRS) libraries into libraries of pre-computed localised models (experts) that can be retrieved and combined at prediction time. RESULTS: The liblex algorithmic framework was evaluated in a challenging cross-continental test case. A large and heterogeneous DRS library of North American soils was used to construct a library of predictive models for total carbon (TC), which were then applied to samples from a region in the Democratic Republic of the Congo. In this scenario, liblex achieved high predictive accuracy and produced RMSE values that were competitive with, and in this case lower than, those obtained with the benchmark methods considered, including LOCAL, Cubist, global PLS variants, and a convolutional neural network with transfer learning. This work also shows how liblex intrinsically provides uncertainty proxies via prediction dispersion and supports interpretability through stored regression coefficients and variable-importance profiles. SIGNIFICANCE: The liblex framework addresses several long-standing drawbacks of conventional local modelling methods by replacing per-query model refitting with the retrieval and aggregation of pre-computed localised experts. This design reduces computational burden at prediction time, enables deployment without access to the full reference library, preserves interpretability, and supports robust prediction across heterogeneous domains. The dispersion among retrieved expert predictions also provides an intrinsic, low-cost proxy for sample-specific uncertainty. These features make liblex a scalable, interpretable, and privacy-preserving strategy for operationalising large spectral libraries.