In this study we aim to develop an automated method to detect the severity of AMD from color fundus images. The main contribution consists of a stacking ensemble learning approach, which combines the knowledge of five CNN models in order to perform accurate AMD severity grading. Experimental results show that the ensemble method achieves a classification accuracy of 95.2%, a precision of 95.27, a sensitivity of 95.25%, a specificity of 96.68% and an F1-score of 95.11% which outperforms several existing methods.