Modeling of the Variables that Influence Sugarcane Yield using C5.0 and QUEST Decision Tree Algorithms

Introduction The sugar industry usually gathers huge amounts of information during normal production operations, which is rarely used to study the relative importance of both management and environment on sugarcane yield performance. Yield prediction is a very significant problem of agricultural organizations. Each agronomist wants to know how much yield to expect as soon as possible. The aim of this study was to determine the performance of C5.0 and QUEST algorithms to predict the yield of sugarcane production in Amir-Kabir agro-industry Company of Khuzestan province, Iran. However, the working method described in this paper is applicable to other geographical areas and other kinds of crops. Materials and Methods The data for the study were collected from Amir-Kabir agro-industry Company. The data is obtained from 2012 to 2016 years. The study area is located in Khuzestan Province which is a major agricultural region in Iran. The geographical location of the study area is between latitudes 31° 15′ to 31° 40′ north and longitudes 48° 12′ to 48° 30′ east. It covers an area of about 12000 ha. The average elevation of the study area is 8m above sea level. Mean annual rainfall within the study area is 147.1mm, the mean annual temperature is approximately 25°C and the mean soil temperature at 50cm depth is 21.2°C. The used data were obtained from a survey with 15 variables carried out on 1201 sugarcane farms. Variables used in the study of data mining can be divided into two categories: target variable and predictor variables. The variable of yield was used as the target variable (dependent) and other variables as predictor variables (independent). In two models, the input data included crop cultivar, month of harvest, chemical fertilizer (Nitrogen), chemical fertilizer (Phosphate), age (plant or ratoon), times irrigation, ratio of surface spraying, soil texture, soil electrical conductivity (EC), water consumption per hectare, drain, farm management, crop duration, area, and yield-category. The study was included in 1201 farms. The necessary data were collected and pre-processing was performed. We propose to analyze different decision tree methods (C5.0 and QUEST). Results and Discussion First, decision tree methods were analyzed for variables. Then, according to C5.0 method (error rate 0.2319 for the training set and 0.3306 for test set) performed slightly better than another method in predicting yield. Crop cultivar is found that an important variable for the yield prediction. 24 rules were found in this study, C4.5 showed a better degree of separation. The measured prediction rate of C5.0 was correct: 76.81% and wrong: 23.19% in the training data, and correct: 66.94% and wrong: 33.06% in the test data. The prediction rate of QUEST was correct: 68.25% and wrong: 31.75% in the training data, and correct: 70.83% and wrong: 29.17% in the test data. Using the training data comparison between the model types showed that the C5.0 model produces a more accurate prediction model and was, therefore, the model to use. Using the testing data in comparison with the model types showed that the QUEST model produced a more accurate prediction model. The results of our assessment showed that C5.0 and QUEST algorithms were capable to produce rules for sugarcane yield. Therefore, our proposed methods as an expert and intelligent system had an impressive impact on sugarcane yield prediction. Conclusions In today's conditions, agricultural enterprises are capable of generating and collect large amounts of data. Growth of data size requires an automated method to extract necessary data. By applying data mining technique it is possible to extract useful knowledge and trends. Knowledge gained in this manner may be applied to increase work efficiency and improve decision making quality. Data mining techniques are directed towards finding those schemes of work in data which are valuable and interesting for crop management. In this research, decision tree algorithms (C5.0 and QUEST) were used. This classification algorithm was selected because it has the potential to yield good results in prediction and classification applications. This study was performed to present a model-based data mining to predict sugarcane yield in 2012-2016. The 24 classification rules generated from the C5.0 decision tree algorithm have great practical value in agricultural applications. The results showed the QUEST and C5.0 decision tree algorithms produced the best prediction accuracy. Sensitivity analysis results indicated that crop cultivar was the most important variables. It was observed that efficient technique can be developed and analyzed using the appropriate data, which was collected from Khuzestan province to solve complex agricultural problems using data mining techniques (decision tree). The decision tree has been found useful in classification and prediction modeling due to the fact that it can capability to accurately discover hidden relationships between variables, it is capable of removing insignificant attributes within a dataset.