Performance of Multilayer Perceptron Neural Network Models and Radial-Based Functions in Estimation of Sugar-cane Crop Yield

According to the high importance of sustainable crop production in the agro-industry units, intelligent systems such as artificial neural networks should be used to manage farm units.Therefore, the main purpose of this study was to compare the performance of MLP (Multi-Layer Perceptron) and RBF (Radial Basis Functions) neural network models in order to modeling and estimating of the sugarcane crop yield and investigate the factors affecting it.  

The study was analytical and its database contained of a matrix  elements. Required data for this research were obtained from the Debel Khazaei sugar cane agro-industry farm during the years 2016 to 2019. The input variables and their units were soil electrical conductivity (dS.m-1), Phosphate and Nitrogen chemical fertilizer (kg.ha-1), water consumption (m3.ha-1), also, irrigation times, month of harvest, age of crop, sugarcane variety, soil texture (non-dimensional), respectively. The analysis was performed by MATLAB 2017 software.  

By comparing the error parameters of RMSE (Root Mean Square Error) and the MAPE (Mean Absolute Percentage Error), and according to indexes of R2 (coefficient of determination) and the EF (Model Efficiency) and, in the validation phase the RBF model was the best model with 0.064494 (%), 0.037686, 0.7576 and 0.800409 (non-dimensional) respectively. Also, the RBF model indicated that the sugarcane variety and soil electrical conductivity were the most important factors affecting the sugar-cane yield.  

By selecting the appropriate variety of sugarcane and controlling the amount of electrical conductivity of the soil, the yield per unit area can be increased, resulting in greater productivity of the inputs and more sustainable production.