Prediction of Irrigated Wheat Yield by using Hybrid Algorithm Methods of Artificial Neural Networks and Genetic Algorithm

In recent decades, the most important issue for agricultural activities is maximizing the productions. Today, wheat is grown on more lands than any other commercial crops and continues to be the most important food grain source for humans. Sustainable agriculture is a scientific activity based on ecological principles with focus on achieving sustainable production. It requires a full understanding of the relationships between crop production with soil and land characteristics. Furthermore, one of the objectives of sustainable agriculture is enhancing the agricultural production efficiency through applying proper management, which requires a deep understanding of relationships between production rate, soil and environment characteristics. Hence, the first step in this process is finding appropriate methods which are able to determine the correct relationships between measured characteristics of soil and environment with performance rate. The aim of this study was evaluating the performance of neuro-genetic hybrid model in predicting wheat yield by using land characteristics in the west of Herris City.
The study area was located in the northwest of east Azarbaijan province, Heris region. In this study, 80 soil profiles were surveyed in irrigated wheat farms and soil samples were taken from each genetic horizon for physical and chemical analyses. In this region, soil moisture and temperature regimes are Aridic border to Xeric and Mesic, respectively. The soils were classified as Entisols and Aridisols. We used 1×1 m woody square plots in each profile to determine the amounts of yield. Because of nonlinear trend of yield, a nonlinear algorithm hybrid technique (neural-genetics) was used for modeling. At first step, the average weight of soil characteristics (from depth of 100 cm) and landscape parameters of selected profiles were measured for modeling according to the annual growing season of wheat. Then, land components and wheat yield were considered as inputs and output of model, respectively. For this reason, genetic algorithm was investigated to train neural network. Finally, estimated wheat yield was obtained using input data. Root mean square error (RMSE) and Coefficient of determination (r2), Nash-Sutcliffe Coefficient (NES) indices were used for assessing the method performance.
The sensitivity analysis of model showed that soil and land parameters such as total nitrogen, available phosphorus, slope percentage, content of gravel, soil reaction and organic matter percentage played an important role in determining wheat yield in the studied area. The soil organic matter and total nitrogen had the highest and lowest correlation with wheat yield quantity and quality, respectively, indicating the total nitrogen was the most important soil property for determination of wheat yield in our studied area. We found that network learning process based on genetic algorithms in the learning process had lower error. The findings showed that beside of confirming the desired results in the case of using sigmoid activation function in the hidden layer and linear activation function in the output layer of all neural networks, it is demonstrated that the proposed hybrid technique had much better results. These findings also confirm better prediction ability of neural network based on error back propagation algorithm or Levenberg-Marquardt training algorithm compared to other types of neural network confirms.
Using nonlinear techniques in modeling and forecasting wheat yield due to its nonlinear trend and influencing variables is inevitable. Recently, genetic algorithms and neural network techniques is considered as the most important tools to model nonlinear and complex processes. Despite the advantages of these techniques there are a lot of weaknesses. Imposing specific conditioned form by researchers in the techniques of genetic algorithms and stopping neural network learning at the optimal points are the main weaknesses of these techniques, while searching for global optimal point and not imposing a specific functional forms are the robustness of genetic algorithm techniques and neural networks, respectively. Results of this study indicated that the proposed hybrid technique had much better results. Correlation coefficient (0.87) and average deviation square error (473.5) were high and low, respectively. It can be concluded that the surveyed soil properties have very strong relationship with the yield. Implementation of appropriate land management practices is thus necessary for improving soil and land characteristics to maintain high yield, preventing land degradation and preserving it for future generations required for sustainable development.