What is the Best Inference Trajectory for Mapping Soil Functions: An Example of Mapping Soil Available Water Capacity over Languedoc Roussillon (France)

Abstract

Extending digital soil mapping to the mapping of soil functions that can support end-user decisions comes to coupling a digital soil mapping procedure and a soil function assessment method. This can be done following various possible inference trajectories following the order with which “combining primary soil properties”, “aggregating soil layers across depths” and “mapping” are executed to provide the targeted output. Eighteen inference trajectories, designed for computing soil available water capacity maps in the Languedoc–Roussillon region (France), were compared with regard to their mapping performances. The best performance (SSMSE = 0.42) was obtained by a trajectory that, before mapping, combined the three first GlobalSoilMap soil layers and computed the available water capacity of each layer. The worst (SSMSE = 0.07) was observed when all the soil layers and soil properties were combined prior to mapping. We explain the observed differences between trajectories by examining the differences in mapping errors and in error propagation between the compared trajectories, which involve both the correlations between the soil properties and between their mapping errors. This paves the way to spatial soil inference systems that could perform an ex ante selection of the best possible inference trajectory for mapping a soil function.

Estimating soil organic carbon stock change at multiple scales using machine learning and multivariate geostatistics

“…We used a direct mapping approach for modelling the spatial distribution of SOC stocks in 1992 and 2010, and then derived the SOC stock change between the two years since the cokriging approach yields coherent results as we pointed out in Section 4.2. However, there are further alternative trajectories or approaches (Heuvelink and Pebesma, 1999;Styc and Lagacherie, 2019) that may be interesting to check and compare with the approach used in this study, since Styc and Lagacherie (2019) observed important differences of performances between such inference trajectories.…”

Effects of climate change on forest plantation productivity in Chile

“…The soil depth map was generated using a method based on survival analysis (Chen et al, 2019; Ishwaran et al, 2008; Styc & Lagacherie, 2019) and a random forest model (Ishwaran, 2020) determining the most probable value for soil depth with a 90% confidence interval. Comparison between the predicted and observed soil depth map produced R 2 = .67 and RMSE = 0.67 m.…”

Available water capacity from a multidisciplinary and multiscale viewpoint. A review

“…DSM has been increasing applied to AWC in recent years (Hong et al 2013;Jin et al 2018;Levi et al 2015;Roman Dobarco et al 2019a;Szabo et al 2019;Zare et al 2018), and several articles include pioneer considerations of rooting depth and gravel contents (Leenaars et al 2018). Styc and Lagacherie (2019) showed that performances of AWC mapping were influenced strongly by the order in which "combining primary soil properties," "aggregating soil layers across depths," and "mapping" are executed to provide the targeted AWC.…”