Predicting soil phosphorus and studying the effect of texture on the prediction accuracy using machine learning combined with near-infrared spectroscopy.

The estimation of soil phosphorus is essential for agricultural activity. The laboratory chemical analysis techniques are expensive and labor-intensive. In the last decade, near-infrared spectroscopy has been become used as an alternative for soil attributes analysis. It is a rapid technique, and inexpensive relatively. However, this technique requires a calibration step using different machine learning and chemometrics tools. This study aims to develop predictive models for total soil phosphorus and extractable phosphorus by the Olson method (P-Olson) using three regression methods, namely partial least squares (PLS), regression support vector machine (RSVM) and backward propagation neural network (BPNN), combined with a proposed variable selection algorithm (PARtest) and a genetic algorithm PLS (GA-PAS). Also, it aims to investigate the effect of the texture on the accuracy of the prediction. The results show that PARtest combined with PBNN outperform the other used algorithms with an R = 0.86, RMSE = 1104 mg kg, and RPD = 3.23 for the TP. For P-Olson the RSVM coupled with GA-PLS outperforms all other methods with an R = 0.77, RMSE = 20.09 mg kg, and RPD = 1.90. The use of hierarchical ascendant clustering (HAC) helps to reduce the heterogeneity of soil and helps to increase the quality of prediction. The obtained results show that the models for clayey and loamy soils yielded an excellent prediction quality with an Rt = 0.88, RMSEt = 857.33 mg kg, and RPD = 4.10 using BPNN with PARtest for TP. Furthermore, an R = 0.83 RMSE = 8.30 mg kg, RPD = 11.00 3.11using RSVM with GA-PLS for P-Olson. Thus, the texture has a significant effect on the prediction accuracy.