Diabetic retinopathy (DR), one of the common chronic complications of diabetes, is of great significance in clinical practice as its accurate classification helps ophthalmologists tailor treatment plans for patients. Diabetic macular edema (DME), a complication closely related to DR, is often used for multi-task learning with DR to assist in DR diagnosis. Currently, deep learning methods for DR grading diagnosis mainly focus on network architecture design, while research on data augmentation techniques is relatively limited. This paper proposes a novel data augmentation method, GreenBen, that combines feature redundancy reduction of the green channel with the background suppression capability of Ben enhancement. Despite its simple design, this method is highly effective. Extensive experiments conducted on three public datasets show that GreenBen achieves stable and significant performance improvements compared with other data augmentation methods, with an average accuracy improvement of 4% and a maximum improvement of up to 10%, regardless of whether it is used for single DR classification or multi-task joint classification, or with CNN or Transformer models.