Creative Commons source reading · Data quality and algorithm evaluation
Improving corrosion depth estimation in buried pipelines using SMOTE-integrated machine and deep learning models
Mohamed Khalid AlOmar; Mohammed Majeed Hameed; Osama Khamees Ali; Aleema Ahmad Samir; Tanya Kukreja; Adil Masood; Ali Salem
Originally published in PLoS One · 2026-08-27
From the article
Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.
Abstract
Accurate prediction of corrosion depth in buried pipelines is essential for integrity assessment, reducing inspection and maintenance costs, and managing risk in pipeline transportation systems. Although machine learning (ML) and deep learning (DL) techniques have been widely applied to corrosion prediction, their reliability depends strongly on the quality, representativeness, and balance of the training data. In many previous studies, data imbalance has remained a key limitation, leading to biased models that mainly learn the majority class, underestimate severe corrosion cases, and exhibit reduced overall performance. In this study, a polynomial-based Synthetic Minority Over-sampling Technique (SMOTE) is employed as a data-level solution to alleviate class imbalance and improve model learning. SMOTE is integrated with several…
4. Results of the models
The current paper investigates the effectiveness of DLNN and other ML models (SVR, XGBoost, and MLR) in predicting the maximum depth of corrosion in pipes (dmax). In this work, to enhance prediction accuracy and improve model performance, we used synthetic data to increase the minority rates in the dataset by applying the SMOTE technique. This approach generates synthetic minority class samples to address data imbalance, leading to better performance and more reliable predictions. To obtain a fair assessment, the models were first trained without applying the SMOTE technique, and their hyperparameters were tuned accordingly. In the next step, the same model structures and the same optimized hyperparameters were used to train the models with SMOTE-generated synthetic data. This procedure allows for an accurate evaluation of the…
4.2. Application results
The performance of the applied models in predicting dmax was evaluated using several statistical metrics before and after applying the SMOTE technique. Initially, without SMOTE, the models generally showed good performance during the training phase, as reported in Table 3. XGBoost and DLNN exhibited almost remarkable accuracy, achieving low RMSE values of 0.650 and 0.659, respectively, and XGBoost slightly outperformed DLNN with MAE = 0.387 and MAPE = 28.74%, compared with MAE = 0.412 and MAPE = 34.50% for DLNN. The SVR model delivered lower prediction accuracy than both DLNN and XGBoost, with higher error values (MAE = 0.514, RMSE = 0.568, MAPE = 43.52%), while the MLR model provided the weakest performance, showing the highest prediction errors. After implementing SMOTE, all AI models demonstrated significant improvements in the…
Editorial reading guide
- What data created and tested the model, and was evaluation independent?
- What is the cost of a false positive, false negative or out-of-distribution decision?
