Towards Reliable Data Augmentation in Machine Learning

This repository contains the data and code for the paper "Towards Reliable Data Augmentation in Machine Learning: Practices to Prevent Data Leakage." It includes two case studies demonstrating full reproducibility (Tomato Disease Detection and Rice Pest Detection), which show that applying data augmentation before the train-test split leads to severe data leakage and massively inflated performance metrics. These resources enable the complete replication of our findings. Highlights of the paper: - Data leakage from pre-split augmentation inflates mAP50 by up to 66 percentage points. - Two forms of leakage are identified: process-induced and data-inherent data leakage. - Leakage is a pervasive systemic problem across applied ML domains. - Actionable best practices are provided for researchers, reviewers, and educators.