Accurate crop recommendation systems are essential for optimizing agriculturalproductivity and sustainability, yet existing approaches often fail to integrate diverse environmental factors and adapt to location-specific conditions. This study proposes a hybrid machine learning model that leverages soil and climate parameters through a threestage pipeline: Random Forest for feature selection, Extreme Gradient Boosting for robust prediction, and a lightweight Feed forward Neural Network for final decision-making. The model was evaluated on real-world datasets, demonstrating superior performance with an overall accuracy of 95.3%, precision of 95.1%, recall of 95.0%, F1-score of 95.1%, and a root mean squared error (RMSE) of 0.12. Ablation experiments reveal that excluding Random Forest feature selection reduces accuracy to 91.9%, omitting geospatial adaptation lowers accuracy to 92.6%, and replacing the neural network with logistic regression drops accuracy further to 89.2%. These results confirm the effectiveness of the hybrid architecture and the critical role of feature selection and geospatial adaptation in enhancing crop recommendation accuracy. This work presents a scalable and locationsensitive framework with significant potential to advance precision agriculture.