ENABLING FIELD-BASED PALEOCLIMATE CHARACTERIZATION VIA LASER-INDUCED BREAKDOWN SPECTROSCOPY (LIBS) COUPLED WITH MACHINE LEARNING ALGORITHMS

Climofunctions are ratios of major elements used to calculate past climate conditions when measured from paleosols. Application of climofunctions requires calibration of elemental compositions from a suite of modern soils from the study area with known climate conditions. We evaluate multiple approaches to calibrating climofunctions using laser-induced breakdown spectroscopy (LIBS), a handheld analytical field-capable device which provides instantaneous in-situ acquisition of elemental compositions for solid materials. LIBS operates by capturing atomic emission spectra, where intensities of specific emission peaks reflect the concentrations of corresponding individual elements. For calibration of climofunctions in the central Andes, 107 modern soil samples were collected in southern Peru and western Bolivia as 4 multi-latitude transects covering the full range of elevations and climate conditions present. Mean annual precipitation and temperature used in this calibration were obtained from high-resolution open-source datasets extending over 40 years.

In addition to conventional climofunctions defined above, LIBS was coupled to backwards propagating artificial neural networks (BP-ANNs) to assess a novel approach which directly estimates climate conditions from raw LIBS spectra. ANNs are well suited tools for identifying complex patterns in dense datasets, like LIBS spectra, relationships to desired outputs, such as climate parameters, and constructing predictive models or classifications from these insights. This multivariate technique allows rapid and direct assessment of climate conditions without the necessity to calculate and calibrate distinct climofunctions for each climate parameter. A subset of 20 soil samples representing the full range of climate conditions was set aside for performance testing of two supervised BP-ANN models evaluated with cross validation, R2, and mean absolute percentage error. Preliminary results indicate good ability for ANN-enabled LIBS to assess climate conditions within well-characterized ranges, but poor accuracy in calculating extreme climate conditions. Ongoing work focuses on constructing and evaluating an alternative machine learning approach – clustering algorithms – which classify samples into climate regimes rather than quantifying individual climate parameters.