Introduction

Sugarcane is cultivated on 135,000 hectares of agricultural land in Khuzestan Province. In some of the farms of Sugarcane and By Products Development Company, an average of 125 tons per hectare is harvested annually, which indicates the favorable potential of Khuzestan province for sugarcane cultivation. In recent years, sugarcane production yield per unit area has decreased. Therefore, extensive studies are needed in this field to investigate the causes of the decline in sugarcane yields and provide solutions to improve crop yield. Sugarcane yield is affected by various parameters; given that simultaneous examination of the effects of management, water, soil, and climate on sugarcane plants is not possible in the field method; therefore, modeling is a solution to this problem. Data mining and machine learning tools can analyze data and identify factors affecting product performance. The purpose of this research is to identify problems and estimate the performance of sugarcane farms in Dehkhoda Agro-Industry using the support vector machine and random forest algorithms.

Materials and Methods

To collect data and information, statistics from Sugarcane and By Products Development Company and the Dehkhoda Sugarcane Agro-Industry Company in the 2019-2023 crop years were used, as well as field data collection and measurement. The data used includes 2032 records, which consist of four categories of datasets: meteorological data, crop data, soil data, and farm management data. The variables studied in the two algorithms of support vector machine and random forest include: crop variety, harvest month, nitrogen chemical fertilizer, phosphorus chemical fertilizer, plant age (first crop or ratoon), number of times the field is irrigated during the growing season, ratio of the field's sprayed area (ratio of the sprayed area to the total field area), soil texture, soil electrical conductivity, amount of water consumption per hectare, drainage rate, field management (production management), length of the growing season, field area, sugarcane yield, and climatic variables. The indicators used to evaluate the accuracy of the model included: root mean square error, average absolute value of error, and correlation coefficient. Python software version 3.12.4 was used for modeling. In this study, 70% of the data, which includes 1626 samples, were selected for the training set and 30% of the data, which includes 406 samples, were selected for the testing set.

Results and Discussion

The scores of the support vector machine and random forest models, considering sugarcane yield as the target variable in two training and testing datasets, were 84.5 and 85.6 percent on the testing data, respectively, and 86.9 and 87.1 percent on the training data, respectively. In the experimental phase, the performance of the random forest model was significant with a correlation coefficient (R2) of 0.92, a mean absolute error (MAE) of 0.34, and a root mean square error (RMSE) of 62.81. In this study, various metrics and methods were used to evaluate the accuracy of the models and their stability and reliability in dealing with new data. The constructed models were evaluated using data cross-validation, confusion matrix, performance characteristic curve, and sensitivity accuracy curve. For sugarcane yield prediction, the accuracy of the hybrid classifier with soft voting on training and test data was 93% and 89%, respectively. The accuracy-frequency for the sugarcane yield prediction model using random forest model was 96%, which is a significant achievement in this research. This high accuracy indicates that the model was able to well identify and model the complex and nonlinear patterns present in sugarcane data. Also, this high accuracy indicates that the input variables selected for the model had high predictive power. The obtained random forest output curve shows that the random forest algorithm has also been able to perform acceptable in predicting sugarcane yield. The upward trend and then stabilization of the curve indicates that the model has converged well and reached the optimal point.

Conclusion

The results showed significant success in using machine learning techniques to predict the performance of sugarcane fields. The 97% accuracy of the random forest model, confirmed by 10-layer cross-validation, indicates that this model can be used as a reliable tool in sugarcane plantation management. Complementary results from the vector machine algorithm also confirm these findings. The balanced distribution of data in the box plot, good performance in the confusion matrix, and favorable ROC curve all indicate that these models can play an important role in sugarcane plantation management decisions. These results also show that using the random forest technique in agriculture can help improve productivity and better resource management.