Introduction:Predicting crop yield using data has an important impact on socio-economic issues and political decision-making at the regional scale. Guilan province is considered as the hub of hazelnut production in the country. In the science of statistics, there are various methods for categorizing, recognizing patterns, forecasting and modeling data. To predict the performance of agricultural products with data mining, various methods such as K-nearest neighbor (KNN), Classification and Regression Tree (CART), neural network feedforward backprop (FFB) and multi linear regression (MLR) are leading. Since the prediction of hazelnut yield using data mining methods is very important for the decision makers of the agricultural sector in terms of knowing the positive and negative factors in the yield. This research aims to answer this need by using all these methods for hazelnut orchards equipped with drip irrigation system in Guilan province.
Methodology :The information used in this study includes the amount of water consumption and yield of hazelnut and the parameters related to these two indicators in hazelnut orchards equipped with pressurized irrigation systems under the management of operators in the cities of Guilan province, including Roodsar, Amlash, Siyahkol and Roodbar, which in a research The measured quantities included the salinity of the irrigation water, the salinity of the saturated soil extract, the number of times of irrigation, the yield of corn, and the volume of water consumed per tree in each irrigation round. To determine the characteristics of the soil in each garden, the depth of activity The roots (0-30 cm) were sampled from the soil. At the same time, water samples were also obtained from the gardens of the region and the electrical conductivity of the water samples was measured. The volume of water needed in each round of irrigation for each tree was obtained by measuring the irrigation water. The percentage of sand, silt and clay was also obtained by determining the soil texture by hydrometric method. The amount of yield in one hectare was also obtained after the harvest at the end of the harvest in the year 2021-2022.
Results and Discussion :To model the input variables including daily maximum evaporation and transpiration data, electrical conductivity and soil acidity, percentage of clay and percentage of silt and percentage of sand, electrical conductivity and acidity and irrigation volume for each tree in each irrigation round and hazelnut yield as the output of the model. Work was done. The results obtained from each of the models were compared with the measured values using statistical indices of correlation coefficient (R), root mean square (RMSE), Nash Sutcliffe coefficient (NSE) and the best model was selected. The variables of volume of water required by the tree in the irrigation cycle, irrigation water reaction index, soil silt percentage and soil reaction index showed a positive correlation with crop yield.Also Irrigation hours, electrical conductivity of irrigation water, sand percentage, clay percentage, sand percentage, soil electrical conductivity and maximum daily evaporation and transpiration showed a negative correlation with crop yield. The results showed that the leading artificial neural network model performs better than the other three models due to higher correlation coefficient statistics (0.98) and higher Nash-Sancliffe coefficient (0.96). Also, CART decision tree has correlation coefficient (0.93), Nash Sutcliffe coefficient coefficient (0.93) and K-nearest neighbor method has correlation coefficient (0.9), Nash-Sancliffe coefficient (0.7) and multiple linear regression method. The variable has correlation coefficient (0.67), Nash-Sancliffe coefficient (0.47), which indicates higher accuracy of CART decision tree method
Conclusion: By evaluating the results, it was found that the forward neural network method with the error back propagation method provides more accurate results than the decision tree, K-nearest neighbor, and multivariate linear regression models. The CART decision tree method was more accurate than the K-nearest neighbor method, which is the reason for the higher efficiency of the decision tree model due to the uncertainty in the phenomena related to water and soil or the appropriateness of the measured values. The multivariate linear regression model also cannot provide accurate results due to the non-linearity of water and soil characteristics. Therefore, artificial intelligence techniques can be introduced as an efficient tool for developing correct management plans in the field of product performance.