A Comparative Analysis of Ensemble and Deep Learning Architectures for Millet Yield Prediction

Abstract

South Indian millet cultivation remains highly vulnerable to changes in climatic conditions along with soil factors and irregular patterns of rainfall. A machine learning system is developed to make forecasts of millet yield in South Indian districts through the analysis of agro-climatic conditions alongside soil information and crop growing parameters. The information needed preprocessing to encode data and normalize variables before reducing dimensions through Principal Component Analysis. To improve model accuracy the approach employed interaction terms together with polynomial transformations as feature engineering. The experimented model selection process included FCCN-RF alongside ResNet-XGB combined with ResMLP-XGB-LGB and DNN-XGB-LGB-CatBoost. The ensemble model of DNN-XGB-LGB-CatBoost achieved highest performance with 0.9910 R² after conducting assessment with R², MSE, RMSE and MAE. The findings indicated that ResNet-XGB achieved 0.9542 ,ResMLP-XGB-LGB achieved 0.9441 and FCCN-RF achieved 0.9402. Ensemble machine learning algorithms that facilitate sustainable decision-making to address climate-related risks and improve millet production support high accuracy.