The yield of tropical crops is highly sensitive to climate conditions, and accurately modeling the meteorological-driven mechanisms is crucial for improving tropical agricultural productivity and climate adaptability. This study systematically compared the prediction performance of six machine learning models, including LGBM, RF, XGBoost, AdaBoost, SVM and MLR based on natural rubber, mango, pineapple and banana in Hainan. The SHAP method was used to quantify the contribution and non-linear response characteristics of meteorological factors. The LGBM model demonstrated the best prediction performance, with an average R² of 0.945 for the test set (the R² of rubber, mango, pineapple and banana were 0.942, 0.902, 0.954 and 0.983, respectively), and average RMSE and MAE of 1.436 t/hm² and 1.150 t/hm², significantly outperforming the other models (the R² of RF, XGBoost, AdaBoost, SVM, MLR were 0.773, 0.563, 0.589, 0.368 and 0.508, respectively). The meteorological-driven mechanisms exhibited significant crop-specific differences. Rubber yield was mainly driven by solar radiation (the contribution was 14.7%) and temperature factors (the contribution of monthly minimum temperature and monthly maximum temperature were 14.4% and 11.7%, respectively). Mango yield was highly sensitive to monthly maximum temperature (the contribution was 19.0%) and vapor pressure deficit (the contribution was 18.5%). Pineapple and banana yield were dominated by soil moisture (the contribution was 18.9%) and relative humidity (the contribution was 23.6%), respectively. Based on the findings, differentiated agronomic management recommendations for each crop type were proposed. This study demonstrates that machine learning, combined with explainability methods, can effectively elucidate the climate response mechanisms of tropical crops, providing theoretical support for regional agricultural precision management.