Data augmentation increases the diversity of existing training data by applying various transformations like rotation, scaling, and cropping to existing data and eliminating the need to collect additional data. This helps improve model robustness and reduce overfitting.