Meta-knowledge guided Bayesian optimization framework for robust crop yield estimation

Abstract

Accurate pre-harvest crop yield estimation is vital for agricultural sustainability and economic stability. The existing yield estimating models exhibit deficiencies in insufficient examination of hyperparameters, lack of robustness, restricted transferability of meta-models, and uncertain generalizability when applied to agricultural data. This study presents a novel meta-knowledge-guided framework that leverages three diverse agricultural datasets and explores meta-knowledge transfer in frequent hyperparameter optimization scenarios. The framework’s approach involves base tasks using LightGBM and Bayesian Optimization, which automates hyperparameter optimization by eliminating the need for manual adjustments. Conducted rigorous experiments to analyze the meta-knowledge transformation of RGPE, SGPR, and TransBO algorithms, achieving impressive R2 values (0.8415, 0.9865, 0.9708) using rgpe_prf meta-knowledge transfer on diverse datasets. Furthermore, the framework yielded excellent results for mean squared error (MSE), mean absolute error (MAE), scaled MSE, and scaled MAE. These results emphasize the method’s significance, offering valuable insights for crop yield estimation, benefiting farmers and the agricultural sector.