A yield classification prediction framework based on intelligent agricultural sensor data and meta heuristic optimization ensemble learning

Potatoes, as an important global food crop, currently rely heavily on default parameters or a single search algorithm for model optimization in potato yield prediction research. The performance of gradient boosting models such as XGBoost (eXtreme Gradient Boosting) is highly dependent on hyperparameter configuration, and the grid search method suffers from insufficient spatial exploration, resulting in significant sensitivity to parameters that affect classification results. Therefore, the goal of this study is to enhance the accuracy and stability of yield level classification and to propose an XGBoost model framework with GA (Genetic Algorithm)-PSO (Particle Swarm Optimization) hybrid optimization. This paper constructs an 18-dimensional phenological, hydrothermal, and soil feature set based on measured sensor data from 2015 to 2023 and classifies yields into three levels: low, medium, and high. Subsequently, we use PSO as the backbone to achieve fast parameter convergence and periodically introduce GA genetic crossover and mutation to enhance population diversity, thereby avoiding local optimal traps. The optimization objective function is set to maximize the 5-fold cross-validated macro F1 score. The results show that after GA-PSO optimization, the macro F1 of the model increases to 0.872 ± 0.01, which is 10.66% higher than that of the unoptimized XGBoost (0.788 ± 0.02), and 2.47%, 1.04%, and 1.63% higher than those of grid search (0.851 ± 0.02), single PSO optimization (0.863 ± 0.01), and single GA optimization (0.858 ± 0.01), respectively. The accuracy and Kappa coefficient also reach 0.861 and 0.786, respectively, which are the highest among the compared methods. The SHAP interpretation results further validate that the interaction interval between the temperature suitability index and growing degree days (Idx_T = 0.70–0.85, GDD = 900–1040 °C d) is the key threshold for yield transition. GA-PSO effectively improves model robustness and parameter tuning efficiency through a hybrid mechanism that combines fast convergence and search jumping, providing highly reliable decision support for hierarchical management and water-fertilizer strategy formulation in smart agriculture.