Direct determination of soil texture using laser-induced breakdown spectroscopy and multivariate linear regressions

Abstract

Over recent years, the need for cost-effective methods that allow accurate soil texture analysis has become a major issue. In the agricultural realm, soil texture is an important soil characteristic that drives crop production and field management. However, current laboratory methods used to determine soil texture are expensive and time-consuming, requiring sample preparation prior to analysis. In this paper, we ascertain the use of a commercially available laser-induced breakdown spectroscopy (LIBS) analyzer coupled to multivariate regressions to predict sand, silt, and clay percent in soils. Two strategies were pursued in order to address multicollinearity in the high dimensional LIBS data. The first approach consists of using partial least squares regression (PLSR) on the full spectrum while the second uses a penalized regression method based on the elastic net algorithm. The predictive models were trained 100 times using the repeated k-fold cross-validation method, and were averaged together to estimate the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). Based on an independent dataset (testing), the results showed that 85 to 87% of the sand percent variation was explained by elastic net and PLSR, respectively. Variations of silt and clay were explained up to 80%, using PLSR and elastic net, respectively. Meanwhile, it was found that regardless of the regression method the prediction accuracy was on average lower for the sand fractions (RMSE of 11% and MAE of 8%) and higher for clay (RMSE of 3% and MAE of 1%). Overall, PLSR-based models showed slightly better accuracy than elastic net. Finally, our findings demonstrated the potential of LIBS to determine the texture of a soil.