A robust SMOGN-based framework for season-wise crop yield prediction using outlier-aware regression
Abstract
Agriculture remains a vital sector of the economy, with many people relying on it for livelihood; however, traditional shifting cultivation practices have become increasingly unsustainable due to environmental degradation, soil erosion, and climate vulnerability. Population growth and resource scarcity have further shortened cultivation cycles, reducing productivity and efficiency. This study aims to develop a robust methodology for forecasting crop yields during the growing season by integrating key environmental and agronomic factors to support data-driven crop selection. Various preprocessing steps, including scaling, encoding, outlier detection and treatment, and the Synthetic Minority Over-sampling Technique for Regression with Gaussian Noise (SMOGN), were applied, followed by 5-fold cross-validation to evaluate model performance. Eight regression models, such as k-nearest neighbors (KNNR), support vector regressor (SVMR), random forest regressor (RFR), decision tree regressor (DTR), artificial neural network regressor (ANNR), extreme gradient boosting regressor (XGBoostR), gradient boosting regressor (GBR), and bagging regressor (BR), were compared using MSE, MAE, RMSE, and R². Among them, XGBoostR and RFR consistently outperformed others, with XGBoostR achieving the lowest MSE (1.73119, 2.10232), MAE (0.48531, 0.49618), RMSE (1.17699, 0.90642), and the highest R² (0.90642, 0.88645) after GridSearchCV tuning. For crop-specific sub-datasets, RFR performed best after interquartile range (IQR) outlier treatment and SMOGN oversampling, achieving test scores of MSE = 0.11072, MAE = 0.20065, RMSE = 0.33095, and an average cross-validation R² of 0.96042, while XGBoostR attained the highest overall accuracy (R² = 0.96831) for the annual crop category.