A data-driven spline model designed to predict paleoclimate using paleosol geochemistry

Abstract

Paleosols (fossil soils) are abundant in the sedimentary record and reflect, at least in part, regional paleoclimate. Paleopedology thus offers a great potential for elucidating high resolution, deep-time paleoclimate records. However, many fossil soils did not equilibrate with climate prior to burial and instead dominantly express physical and chemical features reflective of other soil forming factors. Current models that use elemental oxides for climate reconstruction bypass the issue of soil-climate equilibration by restricting datasets to narrow ranges of soil properties, soil-forming environments and mean annual precipitation (MAP) and mean annual temperature (MAT). Here we evaluate a data-driven paleosol-paleoclimate model (PPM_1.0) that uses subsoil geochemistry to test the ability of soils from wide-ranging environments to predict MAP and MAT as a joint response with few initial assumptions.

The PPM_1.0 was developed using a combined partial least squares regression (PLSR) and a nonlinear spline on 685 mineral soil B horizons currently forming under MAP ranging from 130 to 6900 mm and MAT ranging from 0 to 27 °C. The PLSR results on 11 major and minor oxides show that four linear combinations of these oxides (Regressors 1-4), akin to classic oxide ratios, have potential for predicting climate. Regressor 1 correlates with increasing MAP and MAT through Fe oxidation, desilication, base loss and residual enrichment. Regressor 2 correlates with MAT through temperature-dependent dissolution of Na- and K-bearing minerals. Regressor 3 correlates with increasing MAP through decalcification and retention of Si. Regressor 4 correlates with increasing MAP through Mg retention in mafic-rich parent material. The nonlinear spline model fit on Regressors 1 to 4 results in a Root Mean Squared Error (RMSE_MAP) of 228 mm and RMSE_MAT of 2.46 °C. PPM_1.0 model simulations result in Root Mean Squared Predictive Error (RMSPE_MAP) of 512 mm and RMSPE_MAT of 3.98 °C. The RMSE values are lower than some preexisting MAT models and show that subsoil weathering processes operating under a wide range of soil forming factors possess climate prediction potential, which agrees with the state-factor model of soil formation. The nonlinear, multivariate model space of PPM_1.0 more accurately reflects the complex and nonlinear nature of many weathering processes as climate varies. This approach is still limited as it was built using data primarily from the conterminous USA and does not account for effects of diagenesis. Yet, because it is calibrated over a broader range of climatic variable space than previous work, it should have the widest array of potential applications. Furthermore, because it is not dependent on properties that may be poorly preserved in buried paleosols, the PPM_1.0 model is preferable for reconstructing deep time climate transitions. In fact, previous studies may have grossly underestimated paleo-MAP for some paleosols.