Advancing Agricultural Yield Forecasting through Geomagnetic Navigation Algorithm-Optimized Machine Learning: for Accurate Multi-Crop Agricultural Yield Prediction

Accurate crop yield prediction is essential for food security and resource allocation. While Support Vector Regression (SVR) has demonstrated strong predictive performance, its effectiveness depends critically on hyperparameter configuration. Conventional tuning strategies are computationally prohibitive or lack convergence guarantees, and established metaheuristics such as Particle Swarm Optimization (PSO) and Genetic Algorithms (GA) are susceptible to premature convergence. This study proposes GNA-SVR, a framework integrating the Geomagnetic Navigation Algorithm (GNA) with SVR for multi-crop yield prediction. GNA is a swarm intelligence optimizer inspired by the geomagnetic navigation of migratory birds, incorporating three mechanisms: geomagnetic gradient dominance with multi-source cognitive modulation, adaptive cognitive landmark chain correction, and triple heavy-tailed distribution bionic perturbation. The framework is evaluated on five crop types (wheat, maize, rice, soybean, and cotton) using a publicly available agricultural dataset comprising 10{,}478 observations across 101 countries. GNA-SVR achieves a mean R² of 0.957 ± 0.007 and an RMSE of 1,989 ± 74 hg/ha, outperforming six competing methods including PSO-SVR, GA-SVR, WOA-SVR, GWO-SVR, Grid Search, and Random Search. Quantitatively, GNA-SVR delivers a 16.7% reduction in RMSE relative to the closest competitor PSO-SVR (2,387 ± 112 hg/ha), a 20.8% reduction relative to WOA-SVR, a 23.6% reduction relative to GWO-SVR, and an MAE of 1,409 ± 55 hg/ha. The proposed framework reaches 95% of its final fitness within 47 ± 6 iterations with a mean runtime of 38.4 ± 2.1~s per run, which is roughly an order of magnitude faster than Grid Search (312.6 ± 8.2~s). Statistical validation through the Wilcoxon signed-rank and Friedman ranking tests confirms the superiority of GNA-SVR at α = 0.05. Feature importance analysis using SHAP values reveals that average temperature and annual rainfall are the dominant predictors across all crops. These findings establish GNA as a robust hyperparameter optimizer for agricultural prediction tasks.