Modeling soil water content for precision range management

Abstract

I developed site-specific empirical models to predict spring soil water content for two Montana ranches. The models used publicly available Landsat TM 5, USGS DEM, and soil survey-derived data as predictor variables. The goal of the project was to test whether ranchers could collect a limited size soil water content data set, build sitespecific regression models based on the data set, and construct soil water content maps based on the models. The response variable for models consisted of 100 and 82 average soil profile mass water content samples for each ranch, respectively. Half the samples were used for model calibration and half for model validation. Multiple regression models had calibration R2 of 0.64 and 0.43 for each ranch, respectively. Validation showed that the multiple regression models predicted the validation data sets with average error (RMSD) within 0.04 mass water content and regression tree models predicted within 0.055 mass water content. The majority of the validation MSD for all models was accounted for by a lack of correlation between predicted and observed values along a 1:1 line. Models were then constructed with decreased sample sizes. Regression tree models and multiple regression models constructed with 20 to 30 samples predicted soil water content with similar, though still limited, accuracy and precision to full sample models. Site-specific field and lab soil characterization data developed with diffuse reflectance spectroscopy modeling was used to assess the suitability of the soil survey based predictor variables. The multiple regression models with the site-specific data predicted soil water content with average prediction errors (RMSD) of 0.035 and 0.036 mass water content for the two ranches, respectively. Soil survey model predictions were statistically significantly different than site-specific model predictions for one ranch but not the other. Especially dry conditions were a factor contributing to the difficulty in accurately modeling and predicting soil water content encountered at both study sites. Landsat imagery from the peak of the previous growing season, DEM-derived slope and aspect variables, and soil survey attribute data each showed promise as significant predictors of spring soil water content, particularly considering the dry conditions of the data collection period.