A comparative study between a new method and other machine learning algorithms for soil organic carbon and total nitrogen prediction using near infrared spectroscopy

Abstract

Soil organic carbon (SOC) and total nitrogen (TN) are essential elements in agricultural soil and play an important role in many biological and chemical activities for plant growth. The assessment of these parameters in the soil is crucial in agriculture. The problem with traditional chemical analysis methods for SOC and TN is that they are time and resource consuming. In recent years, near-infrared (NIR) spectroscopy has been used as an alternative for SOC and TN determination. Accordingly, in this study, a new approach based on the ensemble learning modelling (ELM) algorithm is used to predict SOC and TN. This approach uses six partial least squares regression (PLS) models with six pre-processing methods as learners for this method. The output of this approach is computed by averaging the predicted values computed by its constituent learners. This algorithm is used to predict the amount of SOC and TN of Moroccan soil collected from four agricultural regions using NIR. The performance of this algorithm is compared with separated regression models, namely PLS, back-propagation neural network (BPNN) with and without variable selection (VS) algorithms. using three metrics, R2, root mean square error (RMSE), and the ratio of performance to deviation (RPD) calculated by a validation dataset. The results show that the ELM outperformed all PLS models and BPNN with and without VS for both elements. Furthermore, BPNN without VS and PLS provided better performance than PBNN with VS for SOC prediction. However, for TN, PLS gave a moderate performance according to other models (R2 = 0.80 and RPD = 2.77). The best predictions were obtained with the E-L model for SOC (R2 = 0.96, RMSE = 1.92, and RPD = 4.87) and TN (R2 = 0.94, RMSE = 0.57, and RPD = 4.91), which classified the model as an excellent one for SOC and TN prediction. The proposed method ELM has the advantage of wider applicability and better performance for TN and SOC quantification by NIR spectroscopy in comparison with separated PLS and BPNN with and without VS.