Probability mapping of soil thickness by random survival forest at a national scale

Soil thickness (ST) is a crucial factor in earth surface modelling and soil storage capacity calculations (e.g., available water capacity and carbon stocks). However, the observed depths recorded in soil information systems for some profiles are often less than the actual ST (i.e., right censored data). The use of such data will negatively affect model and map accuracy, yet few studies have been done to resolve this issue or propose methods to correct for right censored data. Therefore, this work demonstrates how right censored data can be accounted for in the ST modelling of mainland France. We propose the use of Random Survival Forest (RSF) for ST probability mapping within a Digital Soil Mapping framework. Among 2109 sites of the French Soil Monitoring Network, 1089 observed STs were defined as being right censored. Using RSF, the probability of exceeding a given depth was modelled using freely available spatial data representing the main soil-forming factors. Subsequently, the models were extrapolated to the full spatial extent of mainland France. As examples, we produced maps showing the probability of exceeding the thickness of each GlobalSoilMap standard depth: 5, 15, 30, 60, 100, and 200 cm. In addition, a bootstrapping approach was used to assess the 90% confidence intervals. Our results showed that RSF was able to correct for right censored data entries occurring within a given dataset. RSF was more reliable for thin (0.3 m) and thick soils (1 to 2 m), as they performed better (overall accuracy from 0.793 to 0.989) than soils with a thickness between 0.3 and 1 m. This study provides a new approach for modelling right censored soil information. Moreover, RSF can produce probability maps at any depth less than the maximum depth of the calibration data, which is of great value for designing additional sampling campaigns and decision making in geotechnical engineering.