Spatial Simulation of Rice Grain Yields by Empirical Models

Abstract

Rice grain yields can be affected by several parameters including mostly the climate, cultivars, soil types, and the fertilizer managements. In this study, the performance of two empirical models i.e. AquaCrop and artificial neural network models for simulating the grain yields of two most cultivated varieties including the Shirudi (high-yielding) and Tarom (low-yielding) in three contrasting soil series with different fertilizer managements, were calibrated and validated during 2016 and 2017 rice growing seasons for 459 paddy fields in Mazandaran province, northern Iran. Both models were tested by correlation (R2 ), normalized root mean square error (nRMSE) and coefficient of residual mass (CRM). Results indicate that the performance of both models were affected by the rice cultivar, soil type and soil fertilizer management for both calibration and validation datasets. The ability of both models proved to be satisfactorily applicable for spatial simulation of rice yields in northern Iran, but AquaCrop model was superior. The higher accuracy was observed in AquaCrop for the Shirudi cultivar (R2 =0.89; nRMSE=5.87%; CRM=0.03) compared to the Tarom (R2 =0.67; nRMSE=10.41% and CRM=0.04) for 2017 dataset demonstrating that the AquaCrop has more performance for high-yielding cultivar. Two models had lower accuracy in Esmaeilkola soil series with deep clay soil texture. The best accuracy was observed in paddy fields with optimum fertilizer managements for both models. Overall, it is possible to suggest that AquaCrop model could be used to simulate the spatial distribution of rice grain yield with an acceptable accuracy at province-scale and is applicable for local climate change related scenarios.