Feature Selection Techniques for IOT Crop Yield Prediction Using Smart Farming Sensor Data

Abstract

Feature selection plays a critical role in Internet-of-Things (IoT)–based crop-yield prediction due to the presence of heterogeneous, redundant and context-dependent variables derived from soil, climate, management and remote-sensing sources. High-dimensional smart-farming data often degrades generalization performance and increases inference cost, limiting deployment on edge devices. A comprehensive comparative analysis of five feature-selection families: filter, wrapper, embedded, bio-inspired and deep learning–based is conducted using the Smart Farming Sensor Data for Yield Prediction dataset. Fifteen representative methods are evaluated under identical preprocessing, repeated cross-validation and non-parametric significance testing. Embedded SHAP-based selection reduces root mean squared error from 1242.3 to 1186.7 and mean absolute error from 1072.3 to 1030.4 while retaining only 12 features, achieving the strongest accuracy–efficiency trade-off. Bio-inspired multi-strategy whale optimization attains the highest compression, eliminating up to 97.7% of features with competitive RMSE values near 1175 under linear and ensemble regressors. Yield-regime discrimination improves substantially, with distance-correlation filtering and SHAP-select achieving peak AUC–ROC values of 0.571 and 0.560, respectively. Paired Wilcoxon signed-rank tests confirm statistically significant improvements for wrapper and embedded methods (p < 0.05). Results demonstrate that importance-driven embedded selection and multi-objective bio-inspired optimization are well suited for accurate, interpretable and edge-deployable IoT crop-yield analytics.