Soil probiotics are indispensable in agro-ecosystems, enhancing crop yield through nutrient solubilization, pathogen suppression, and soil structure improvement. However, reliable prediction methods for soil probiotics are still lacking. In this study, we use genomic foundation models to generate representations from sample sequences and enhance them by deeply integrating domain-specific engineered features. The enhanced representations enable training a powerful classifier for a target task, rather than relying on conventional parameter fine-tuning. Inspired by the stacking ensemble learning framework, we design a stacked aggregation classifier. It predicts a sample's label by leveraging only a subset of its sequence segments, effectively addressing the challenges in processing long or incompletely assembled sequences. The proposed method is applied to the prediction of soil probiotics and demonstrates excellent performance on both balanced and imbalanced test sets. Furthermore, potential functional genes are revealed from the predicted probiotics, providing valuable biological insights for related studies.