Prediction of soil water infiltration using multiple linear regression and random forest in a dry flood plain, eastern Iran.

• Multiple Linear Regression and Random Forests were used to predict infiltration rate. • Measured soil infiltration ranged from 0.29 to 81.7 mm h−1 with a mean of 13.6 mm h−1. • Channel networks, sand, distance-from-river, NDSI, and elevation were important. • Random Forest best predicted the spatial distribution of infiltration rate. Knowledge of the spatial variation of soil infiltration is necessary for managing water conservation, salinity, and precision agriculture in drylands. In this study, the spatial variation of soil infiltration was investigated using digital soil mapping methods in the Sistan plain, an arid, low-relief flood plain in eastern Iran where a large irrigation project is being implemented to irrigate lands. Information about the spatial variation in soil infiltration will assist the planning of this irrigation project. 138 sampling locations were selected using stratified sampling based on existing polygon-based soil maps. Steady state soil infiltration was measured at each sampling location using the double ring infiltrometer method. Twenty-three environmental covariates were derived from digital elevation models and satellite imagery as well as predictive maps of clay, sand, and silt that were derived from kriging the collected soil samples. A simple (multiple linear regression) and a complex (random forests) model were used to link covariates and infiltration measurements. Ten-fold cross-validation was used to determine model accuracy. Measured soil infiltration ranged from 0.29 to 81.7 mm h−1 with a mean of 13.6 mm h−1. RMSE of the infiltration rate predictions were 13.4 mm h−1 for random forest and 13.9 mm h−1 for multiple linear regression. MAE was 10.5 for random forest and 10.9 for multiple linear regression. The most important covariates were channel networks, sand concentration, normalized difference salinity index (NDSI), and elevation in the random forest model and distance-from-river and sand concentration in the multiple linear regression model. Accuracy metrics for both models were comparable, but the random forest predictions were judged to be closer to reality based on visual review, thus random forests was chosen to make predictive maps of soil infiltration. [ABSTRACT FROM AUTHOR]