<p>Directed energy deposition (DED) is a pivotal additive manufacturing technology, offering substantial advantages for complex and large-scale part fabrication. However, its widespread industrial adoption is hindered by the persistent challenge of in-situ defect formation, particularly porosities, which compromise the quality and reliability of manufactured components. This work addresses the critical challenge of defect detection in DED processes by formulating it as a supervised classification problem. Leveraging a nearly balanced data set and focusing on a single defect type (porosity), a detection framework based on meltpool thermal imaging is developed. The methodology begins with essential preprocessing steps—including region-of-interest extraction and segmentation to isolate meltpool regions—followed by domain expertise-based feature extraction. Two machine learning classifiers—random forest (RF) and support vector machine (SVM)—are employed to identify defect-free and defective thermal images based on the extracted features. Notably, the proposed methodology incorporates the spatio-temporal relationship among thermal images into the modeling, which is normally overlooked in existing literature. The defects of the samples were characterized using computer tomography (CT) imaging to assess the efficacy of the proposed methodology. The experimental results demonstrated the high effectiveness of both classifiers, achieving a high accuracy of detecting defects. However, RF consistently outperformed SVM, attaining an accuracy above 84%. RF also demonstrated superior robustness by reducing false positives, underscoring its reliability for porosity detection in DED processes. This study highlights the advantage of integrating machine learning-based frameworks with domain-specific preprocessing and feature engineering for in-situ defect detection in DED, without relying on expensive and time-consuming characterization methods.</p>