Data augmentation typically refers to the creation of new (training) data/instances (e.g. images) by e.g. modifying existing training/instances data in order to avoid over-fitting and improve the performance (e.g. generalization, precision, recall, etc.) of the model.
There are other techniques that attempt to mitigate over-fitting, for example, regularization techniques (like dropout), but data augmentation changes the data that you use rather than how you use it, or how you define and train/test a model.
When you use data augmentation, you're assuming that more information can be extracted from the original dataset. For example, if you augment your training dataset by rotating existing images of objects, then you assume that this can be beneficial for your task (i.e. you expect your model to encounter rotated versions of the same object).
Typical scenarios where you may need data augmentation are when you have small training or imbalanced datasets.
Padding is not a data augmentation technique because you don't do it to avoid overfitting. Padding is a data preprocessing technique because you do it in order for your model to even be able to process your data.
You could read A survey on Image Data Augmentation for Deep Learning (by Shorten & Khoshgoftaar, 2019, Journal of Big Data), which I have only skimmed through, but it looks like a good paper.
