Improvement in technology and the availability of electroencephalogram (EEG) data have raised the demand for automated seizure detection in long-term EEG recordings. This study proposes a framework to automate seizure detection from long-term EEG by combining anomaly detection, artifact removal, and seizure detection techniques, along with Riemannian manifold and transfer learning approaches. First, the method identifies potential EEG segments for seizures using Riemannian manifold-based features from covariance matrices. Next, it removes extra-physiological artifacts using power-based features. Finally, it uses Riemannian potato-based features to classify the remaining segments with a LightGBM classifier. The method’s performance was evaluated on two datasets–a private dataset (Juntendo) and a public dataset (Siena)–using leave-one-patient-out cross-validation. For the Juntendo dataset, the method achieved an average performance across all subjects with a sensitivity of 89.9%, specificity of 96.8%, precision of 33.3%, and an F1–score of 44.5%. On the Siena dataset, the method achieved a sensitivity of 63.8%, specificity of 98.7%, precision of 32.4%, and an F1–score of 40.5%. Processing EEG data in multiple stages helps reduce the class imbalance problem. Therefore, automating the seizure detection process will ease the practitioner’s workload.