Soil parent material spatial modeling at high resolution from proximal sensing and machine learning: A pilot study

Abstract

Although parent material (PM) is one of the five soil formation factors providing key information on soil variability, the complexity of PM distributions and the difficulty of reaching PM in deep soils prevent its detailed assessment. Proximal sensors such as portable X-ray fluorescence (pXRF) spectrometer and magnetic susceptibility (MS) may be helpful in predicting soil PM in a more practical and accessible way. This pilot study aimed to create spatial PM predictive models for three distinct PMs (charnockite, mudstone, and alluvial sediments) of an experimental farm (Brazil) through random forest (RF) algorithm based on soil samples analyzed via pXRF and MS. Soils were sampled in A and B horizons following a regular-grid design covering the whole study area. The RF algorithm was calibrated to predict PMs using samples from the B horizon of soils with known PM. The prediction model was applied to the area for mapping PM across the whole farm. For validation, PM was identified at 15 different sites and compared with the predicted PM shown on the maps via overall accuracy, Kappa coefficient, producer's and user's accuracies. Al, Fe, Si, Ti, and MS proximal sensor data discriminated well among soils derived from charnockite, mudstone, and alluvial sediments. The map built based on B horizon data showed greater accuracy (overall accuracy = 0.93, Kappa coefficient = 0.85, user's accuracy = 0.92, and producer's accuracy = 0.97) than the map built from the model using A horizon samples (0.73, 0.48, 0.48, and 0.58). These results could represent alternative methods for reducing costs and accelerating the assessment of soil PM spatial variability, supporting soil mapping, and optimized agronomic and environmental decision-making.