Predicting Sorghum Yield in Data-Scarce and Conflict-Affected South Sudan Using Machine Learning Techniques.

Abstract South Sudan continues to experience conflict. Agriculture production remains a crucial food and economic security feature for the overwhelming majority of the population and Sorghum is a key staple crop. In this country, sorghum yield estimation and prediction are of great interest to support interventions from the humanitarian sector, government policies, and for small-scale farmers’ food and economic security. Ongoing conflict, limited access, and poor infrastructure suggest remote sensing and modeling as credible alternative for sorghum yield prediction in this country. This research compares five popular regression techniques, namely Random Forest(RF), Decision Tree, Extreme Gradient Boosting (XGBoost), Support Vector Machine Learning(SVM) and Artificial Neural Network(ANN) Regressor for predicting sorghum crop yield in conflict setting in Upper Nile and Werstern Barh Al Gazal. The study uses 4 years soghum yield data, remotely sensed weather patterns, and above ground biomass proxies including NDVI, EVI, LAI characteristics as inputs for model training and evaluation. Preprocessing method like Fillna is used to handle missing data. The performance of each model is evaluated using metrics like RMSE, MSE, MAE, and R squared. This study wanted to assess the influence of conflict on small-scale farmers’ sorghum yield prediction using remote sensing, climate data, and vegetation proxies in a data-scarce environment. No research so far was done to investigate the impact of conflict on yield prediction in South Sudan. We constructed a modeling framework designed to incorporate data on the likelihood of conflict occurrences sourced from both Uppsala University and Acled, farmers perception on conflict, soil data and remotely sensed vegetation proxies to predict sorghum yield. Machine learning models were used to predict end-of-season sorghum yield and results show that Random Forest model yielded best combination of metrics with an RMSE of 176.27 Kg/ha and an R squared of 58 percent, confirming competitive performance. Outperforming other models, XGBoost records the best RMSE of 171.68 Kg/ha and an R squared of 59 percent. Renowned for its efficacy, XGBoost exceled in capturing intricate relationships within self-declared sorghum yield data, indicating its potential for accurate predictions. Both SVM and ANN achieved relatively lower performance metrics, with an RMSE of 188.95 Kg/ha and an R squared of 50.6 percent for SVM and for ANN a RMSE of 225kg/ha and an R squared of 52.5 percent. These findings for Decision trees, Random Forest, and XGBoost are significant in predicting sorghum yield in South Sudan. The variable importance reveals conflict probability was not a significant factor and did not influence sorghum yield prediction. Cultivated land size was the most significant predictor of sorghum yield in the context of South Sudan. These findings provide insights into the potential efficacy and limitations of machine learning for predicting sorghum yield in conflict settings in support to agricultural planning and humanitarian relief interventions