Assessment of machine-learning methods for the prediction of STN using multi-source data in Fuzhou city, China

Abstract

This study evaluated the performance of machine-learning approaches to predict Soil Total Nitrogen (STN) using remote sensing and environmental data in the coastal city of Fuzhou, Fujian Province, China. Multisource environmental data was combined to identify important variables for topsoil STN distribution prediction. Additionally, STN content was assessed based on environmental covariates. The results from this study showed that random forest (RF), support vector machine (SVM), artificial neural network (ANN), multi-linear regression (MLR), and locally weighted regression (LWR) can achieve high R2 values of 0.96, 0.92, 0.80, 0.97, and 0.93 with respective RMSECV values of 0.08, 0.35, 0.37, 0.43, and 0.65, respectively. Random Forest (RF) was the most effective model among these methods, with the corrosponding highest R2 and lowest RMSECV. RF and SVM models were used to select important predictors; accordingly, RF selected mainly vegetation indexes while SVM selected Visible-Near-Infrared (VIS-NIR) spectra of the soil. Additionally, STN contents had relationships with most environmental covariates derived from remote sensing, soil spectra, and topographic variables. Spectral transformations improved the correlations with STN where the second derivative and standard normal variate transformations produced the best results. This study suggests that machine-learning methods are practical approaches for the prediction of STN and can be used in similar complex coastal environments.