Strategies for predictive digital soil mapping by geophysical, remote sensing and machine learning approaches

Pedology, the study of pedogenesis, includes soil classification and mapping. Digital soil mapping (DSM) has evolved from traditional methods to creating comprehensive spatial soil information systems. This advancement is achieved by integrating field and laboratory data with environmental covariates and incorporating new geotechnologies such as geophysical techniques and remote sensing data, alongside machine learning approaches. This integration in DSM provides novel insights into soil survey and mapping, offering detailed information on soil variability both vertically and laterally. It also raises new research questions that traditional pedology may not have addressed. In this study, we proposed and compared three strategies for DSM in Brazil, creating predictive pedological mapping. These strategies integrate data from three geophysical sensors, remote sensing data, relief, and lithology as input in a machine learning approach testing five algorithms. The four proposed strategies were: i) the combined use of geophysical variables and remote sensing data (G + RS + DEM); ii) the use of remote sensing data only (RS + DEM); iii) the use of geophysical variables only (G + DEM) and; iv) relief data (DEM). Lithology and relief were used as common input data in the predictive pedological mapping modeling process for all four strategies. We conducted a statistical analysis to evaluate the models' performance employing the Kruskal-Walli's test, the F1-score, Kappa, Accuracy, Sensitivity, and Specificity. Additionally, the best strategy was chosen based on the Kruskal-Walli's test and Overall Agreement and Disagreement statistical validation method, utilizing the reference map generated by an expert pedologist. Results revealed that the Random Forest algorithm presented the best performance for modeling predictive pedological mapping in all proposed strategies. Among the predictor variables, the Synthetic Soil Image (a synthetic multi-temporal soil image created by selecting and integrating bare soil observations from satellite data to capture key soil properties for mapping and analysis), relief, and geophysical data had the most significant contributions. While variables associated with remote sensing displayed stronger correlations with surface pedological attributes, geophysical variables demonstrated stronger associations with subsurface pedological attributes and diagnostic horizons. The most effective strategies for predictive digital pedological mapping were the G + RS, while the least effective was DEM. The individual performances of G and RS were comparable. The final predictive digital pedological map had a strong correlation with the traditional one, considering the Agreement/Disagreement validation method. The most significant prediction errors occurred in the transitional zones between pedological and lithological classes. Within the predicted classes, the most substantial errors were observed in classes exhibiting similar morphological attributes such as texture, color, and oxic mineralogy. • Combination of geophysical techniques and remote sensing data to perform soil mapping. • Machine learning applied on proximal and remote sensing data efficiently mapped soil types. • Use of empirical and statistical validation methods in predictive pedological mapping. • Lithological and relief data as input data on predictive models. • Greater importance of relief in relation to geophysics in modeling of pedological classes.