Predicting soil permanganate oxidizable carbon (POXC) by coupling DRIFT spectroscopy and artificial neural networks (ANN)

Infrared spectroscopy has transformed soil property quantification by enabling low-cost, high-throughput analysis of soils, enabling mapping and monitoring of this non-renewable resource. However, less evaluated are newly emerging indicators of soil health. Furthermore, as soil spectral libraries expand in size, commonly employed linear models such as partial least squares regression (PLSR) may be challenged by the number and diversity of spectra. Artificial neural networks (ANN) are an emerging deep learning approach that can offer advantages in quantification of soil properties by utilizing non-linear relationships among spectra and soil components. We compared ANN versus PLSR models for predicting an increasingly used soil health indicator, permanganate oxidizable C (POXC), as well as more routinely predicted soil variables (e.g., clay, soil organic C [SOC]), across a gradient of soil organic matter furnished by a deforestation chronosequence in Kenya (n = 144). Candidate ANN architectures were first methodologically evaluated and described to identify best-practices for the application of ANN to soil spectroscopy. Predictions by the resulting ANN relative to PLSR were similar or slightly improved for routinely measured variables that represent soil organic matter (SOC, C:N) and physical properties (clay, silt, sand, bulk density). The accuracy of POXC predictions were similar for ANN (RMSE 102 mg kg⁻¹) and PLSR (RMSE 106 mg kg⁻¹). However, models drew on shared but also distinct wavenumbers, indicating differential use of information in soil infrared spectra by non-linear versus linear chemometric models. Even in relatively small spectral datasets of similar soil types expected to favor PLSR, ANN shows comparable predictive performance. To help guide future applications of ANN in soil spectroscopy, we propose a systematic procedure to select ANN model hyperparameters.