Predicting fertilizer concentration for precision irrigation under mixed variable-rate fertigation using machine learning: a case study of combined fertilization with dipotassium hydrogen phosphate and potassium chloride

Abstract Variable precise fertigation is critical to precision irrigation. The question of how to monitor the combination of fertilizer concentration and variable irrigation components as accurately as possible is challenging. The primary goal of this study is to develop accurate prediction models integrated with machine learning (ML) to predict the concentration of each type of fertilizers in mixed variable-rate fertigation used for precision irrigation. First, the feasibility of predicting of fertilizer concentration by monitoring physical parameters such as electrical conductivity (EC), acidity (pH) and temperature in mixed variable-rate fertigation was confirmed. 11 selected ML algorithms were applied to develop regression models that can accurately predict each fertilizer concentration of the mixed fertilizer compared to the classical multivariate linear regression (MLR). In addition, cubic spline interpolation (CSI) was used to densify the data sets, and K-fold cross-validation was employed to fairly evaluate the generalization ability ( GA ) of these models. The statistical and diagnostic analyzes revealed the superiority of ML especially SVM, KNN, ETs, and MLP over MLR in predicting each type of fertilizer concentration in mixed variable-rate fertigation with an R 2 range of 0.9499 ~ 0.9970 and an RMSE range of 0.0852 ~ 0.4434 g/L, better than MLR with an R 2 range of 0.8544 ~ 0.9425 and an RMSE range of 0.3752 ~ 0.7559 g/L. Moreover, the contribution of CSI to the modeling accuracy was confirmed, but the sensitivity of the models to EC and pH increased with the data from CSI and the tuning of the model hyper-parameter. Overall, the feasibility and performance of the ML models for predicting mixed fertilizer concentration by monitoring temperature, EC, and pH indicate that the presented ML models have significant application potential for irrigation and fertilization monitoring management of mixed variable-rate fertigation in precision irrigation with high-precision sensor technology.