SoilType: An R package to interplay soil characterization in plant science

Abstract

AbstractYield is a complex quantitative trait whose expression is sensitive to environmental stimuli. Therefore, soil‐related information can increase the predictive ability of genotype's performances across different locations. However, soil information is not always readily available worldwide or before the site or plot level growing season. Thus, in the current version, this tool has two functions. The first function retrieves soil samples and soil (from WoSIS Soil Profile Database) near your target location. Then it predicts 13 soil characteristics (physical and chemical). If the number of samples per location is greater than five, the function uses random forest to predict soil characteristics otherwise it averages the information. The output is a table with the target location and its latitude and longitude coordinates—the number of samples used, the root mean square error (RMSE), and the R‐square for each prediction. From a couple of instances in a trial (location), the second function predicts soil characteristics at the plot level via random forest. The output is a table with the plot IDs, coordinates, number of samples used to make predictions, the RMSE and R‐square for each prediction, the trait predicted, and the predicted value. As a proof‐of‐concept, we used the first function in the LSU Rice Breeding multi‐environmental trials (24 locations), identified the most important soil covariates to determine rice yield, and then clustered the locations. This tool can support breeders in better allocating trials in advance, borrow information from other regions, identify the best variety for each location, reduce costs, and increase accuracy.