Didier Michot

High‐resolution mapping of soil phosphorus concentration in agricultural landscapes with readily available or detailed survey data

Summary High-resolution mapping of soil phosphorus (P) concentration is necessary to identify critical source areas reliably where a large risk of transport coincides with a large potential source of P in agricultural landscapes. However, dense soil P data are not usually available to produce such maps and to obtain them is expensive. In this study, we modelled and mapped soil extractable P (ExtP) and total P (TP) concentrations in an intensively farmed 12-km2 catchment in Brittany (NW France) with two different datasets to test the suitability of readily available regional or national databases for high-resolution mapping. We used a machine learning tool (Cubist) to develop rule-based predictive models from a calibration dataset. Covariates included pedological, geological, agricultural, terrain and geophysical-related attributes obtained specifically in the study area (SURVEY) or derived from readily available regional or national databases (DATABASE). Even though better predictions were obtained with the SURVEY data (RMSE = 0.018 g kg−1 for ExtP and RMSE = 0.219 g kg−1 for TP), the DATABASE data produced acceptable predictions (RMSE = 0.024 g kg−1 for ExtP and RMSE = 0.253 g kg−1 for TP). The machine learning tool helped to identify key covariates that would improve the prediction of soil P when detailed data are not available. Readily available data about crop rotations could increase the accuracy of existing ExtP maps. These maps, combined with additional soil analysis for extractable Al, would improve the mapping of TP and the identification of areas with a large potential source of P. Highlights Modelling and mapping of soil phosphorus with the machine learning algorithm Cubist. Comparison of regional or national databases and detailed survey data for prediction. Models with regional and national data performed well, but some areas with large concentrations of P were not identified. Information about crop rotation and soil extractable Al improved model performance.

Seasonal monitoring of soil salinity by electromagnetic conductivity in irrigated sandy soils from a Saharan oasis

Monitoring soil salinity over time is a crucial issue in Saharan oases to anticipate salinisation related to insufficient irrigation management. This project tested the ability of electromagnetic conductivity surveys to describe, by means of regression-tree inference models, spatiotemporal changes in soil salinity at different depths within a complex 10-ha pattern of irrigated plots in an Algerian oasis. Soils were sandy Aridic Salic Solonchaks with a fluctuating saline watertable at less than 2 m. Apparent electrical conductivity (ECa) was measured by an EM38 device at fixed 10- or 20-m intervals (2889 points) at four sampling dates between March 2009 and November 2010. For calibration and validation purposes, soil salinity was measured from a 1 : 5 diluted extract (EC1:5) in three layers (0–10, 10–25, 25–50 cm) at 30 of these points randomly chosen at each date. ECa measurements were used to predict EC1:5 using calibration regression trees created with the software Cubist, including either parameters specific to the study site (specific model) or more general parameters (general model), allowing extrapolation to other sites. Performance of regression tree predictions was compared with predictions derived from a multiple linear regression (MLR) model adjusted for each date using the software ESAP. Salinity was better predicted by Cubist regression tree models than MLR models. For the deep layer (25–50 cm), Cubist models were more accurate with the specific model (r2 = 0.8, RMSE = 1.6 dS/m) than the general model (r2 = 0.4, RMSE = 2.5 dS/m). Prediction accuracy of both models decreased from the bottom to the top of the soil profile. Salinity maps showed high inter-plot variability, which was captured better by the more flexible regression-tree inference models than the classic MLR models, but they need to build site-specific prediction models. Overall, the monitoring surveys, combined with the Cubist prediction tool, revealed both the seasonal dynamics and spatial variability of salinity at different depths.