Indirect Prediction Based on Machine Learning and Remote Sensing of Ecological Stoichiometric Ratio Superior to Direct Prediction

Abstract

Exploring carbon (C), nitrogen (N), and phosphorus (P) contents as well as dynamic balances in soil are important for understanding ecological characteristics and stability. However, the substantial costs associated with soil surveys limited the possibility of large-scale surveys. The accurate predictive capability of machine learning (ML) supported this possibility. In this study, ML models (Random Forest, RF; support vector machine, SVM; Extreme Gradient Boosting, XGboost; Gradient Boosting Decision Tree, GBDT) and remote sensing data were used to predict soil C, N, and P as well as ecological stoichiometric ratio (ESR) in the Yellow River Delta (YRD). The purpose of this study was to assess the performance of four MLs in predicting soil C, N, and P as well as ESRs and to assess the performance of indirect and direct predictions in ESRs. The results showed that RF and SVM have higher accuracy than XGboost and GBDT. The model was the main factor affecting accuracy, and there were differences in the applicability of different elements to the model. The ESR prediction performance was weaker than that of total elements due to the fact that ESR is controlled by two elements. In the localized prediction of farmland and wetland vegetation, the performance of the models was substantially enhanced compared to the global prediction. The predictive performance of total elements was higher and the predictive performance of ESRs was poorer in soil in farmland. However, this pattern was reversed in wetland vegetated soil. The prediction followed by calculation method improved the prediction accuracy of ESRs. Although not generalizable, this approach still offered a possibility for accurate prediction of multi-element variables such as ESRs. Land-use type had a significant effect on soil C, N, and P as well as ESRs. The mean values of TC, TN, and TP in the study area were 17.730 ± 2.395, 0.710 ± 0.253, and 0.691 ± 0.089 g/kg, respectively. The highest TC was found in farmland and wetland vegetation soils with mean values of 18.228 and 18.138 g/kg, respectively. The Yellow River as well as its old channel had a significant effect on the spatial distribution of C, N, P, and ESRs. This study clarified the spatial distribution pattern of soil C, N, and P as well as ESR in the YRD. In addition, this study provided an indirect prediction method for the prediction of multi-element variables that improved the prediction accuracy.