Evaluation and comparison of AquaCrop model and intelligent models to predict wheat yield (Case study: Miandoab and Mahabad cities)

Due to population growth and Iran's location in arid and semi-arid regions of the world, the need for water and food has increased and as a result, the pressure on water and soil resources will be more than before. On the other hand, the risk of drying up Lake Urmia, which causes environmental problems in the region, requires macro-water planning for the region and the use of optimal cultivation pattern to deal with water scarcity. Therefore, optimal use of water preserves water resources and increases the quality of products. Today more than ever, increasing the production of strategic crops such as wheat requires the proper use of water resources. The main source of food for the Iranian people is wheat and related products, and any action that increases the yield of wheat due to limited soil resources, especially water resources, is important and necessary at the same time. In recent years, significant advances have been made in modeling product growth and development using mechanical models. Plant growth models are increasingly used in the analysis of agricultural systems and simulate the plant's response to growth factors using mathematical equations. The AquaCrop model is one of the dynamic and user-friendly models developed by the FAO. The AquaCrop model receives information about farm, plant, soil, irrigation and climate, and ultimately predicts important parameters such as crop. Wheat yield simulation allows efficient management and better planning under various environmental inputs such as soil and water. To achieve higher accuracy and less model error, field parameters must be properly calibrated by the model to achieve proper performance. Also, calibration of the model, if not done correctly, causes a high error prediction by the model, which leads to incorrect management, water loss, plant drought and other cases. Therefore, using a model that has accurate and close prediction to the AquaCrop model and requires fewer input parameters is essential, which saves time, reduces costs and eliminates calibration errors. However, this model requires relatively large input parameters and is a time-consuming model in the presence of multiple scenarios.3

In recent years, smart models have been able to show high accuracy and become reliable models. Therefore, in the present study, to solve this problem and develop a model with less input data, using the ANN, SVR and SVR-FFA intelligent models and creating 440 scenarios in 2 farms, the performance of the AquaCrop model was compared.99WestW2 farm is located in Miandoab city and has a yield of 6.588 (ton ha-1) and WestW10 farm is located in Mahabad city and has a yield of 5.05 (ton ha-1).

The results of the model are performed using 5 evaluation criteria of Correlation coefficient, Root mean square error, Nash-Sutcliffe coefficient, Wilmot’s index of agreement and, Mean absolute percentage error. The results of this study showed that for both 99WestW2 and WestW10 farms, the SVR-FFA3 model could have the lowest error rate, so that for the yield index, the RMSE value for the mentioned farms was 0.033 and 0.069 (ton ha-1), respectively. The use of three models SVR, SVR-FFA, and ANN and their comparison with the AquaCrop model to predict wheat yield has been done for the first time in this study. The SVR model was able to show the highest accuracy after the SVR-FFA model. For 99WestW2 farm, it can reduce the error rate to 0.043 (ton ha-1) and for WestW10 farm to 0.077 (ton ha-1) and show good performance. The ANN model, after the SVR model, was able to show acceptable accuracy. The ANN model for 99WestW2 farm was able to reduce the error rate to 0.123 (ton ha-1) and for WestW10 farm to 0.094 (ton ha-1). Finally, the ANN model had a relatively higher error than the SVR-FFA and SVR models, respectively, and showed a relatively lower performance than the two models.

Finally, the intelligent SVR-FFA, SVR and ANN models, despite having the least number of inputs, were able to predict yield values in the shortest time and with the highest accuracy. However, the results showed that the lower the model inputs, the weaker the model prediction. For further studies, it is suggested that the ANN model be combined using the firefly algorithm (MLP-FFA) to increase the accuracy of the ANN model and make more accurate predictions of wheat yield.