Application of inclusive multiple model for the prediction of saffron water footprint

Application of inclusive multiple model for the prediction of saffron water footprint

Applying new approaches in the management of water resources is a vital issue, especially in arid and semi-arid regions. The water footprint is a key index in water management. Therefore, it is necessary to predict its changes for future durations. The soft computing model is one of the most widely used models in predicting and estimating agroclimatic variables. The purpose of this study is to predict the green and blue water footprints of saffron product using the soft computing model. In order to select the most effective variables in prediction water footprints, the individual input was eliminated one by one and the effect of each on the residual mean square error (RMSE) was measured. In the first stage, the Group Method of Data Handling (GMDH) and evolutionary algorithms have been applied. In the next stage, the output of individual models was incorporated into the Inclusive Multiple Model (IMM) as the input variables in order to predict the blue and green water footprints of saffron product in three homogenous agroclimatic regions. Finally, the uncertainty of the model caused by the input and parameters was evaluated. The contributions of this research are introducing optimized GMDH and new ensemble models for predicting BWF, and GWF, uncertainty analysis and investigating effective inputs on the GWF and BWF. The results indicated that the most important variables affecting green and blue water footprints are plant transpiration, evapotranspiration, and yield, since removing these variables significantly increased the RMSE (range=11–25). Among the GMDH models, the best performance belonged to NMRA (Naked Mole Ranked Algorithm) due to the fast convergence and high accuracy of the outputs. In this regard, the IMM has a better performance (FSD=0.76, NSE=0.95, MAE) = 8, PBIAS= 8) than the alternatives due to applying the outputs of several individual models and the lowest uncertainty based on the parameters and inputs of the model (p = 0.98, r = 0.08).