Social media allows the spread of vital information regarding crises and emergencies. Thus, emergency management systems can benefit from social media because they can be used to inform the public to take the appropriate precautions. However, social media is riddled with irrelevant information. Therefore, researchers have recently focused on developing robust event detection (ED) systems to extract relevant events and to define their types by relying on deep learning techniques (DL). Hence, this paper proposes an event detection model that merges the DL approach (e.g., MobileBERT) and a novel feature selection (FS) method to improve performance. MobileBERT is a transformer-based model designed to extract features from a text dataset, while the FS is used to preserve the relevant features and to reduce feature representation space. The developed FS method depends on improving the sparrow search algorithm (SSA) using manta ray foraging optimization (MRFO) operators. The modification is conducted to enhance the exploitation ability of the SSA using the operators of MRFO as a local search method. To validate the proposed framework, experiments are conducted using real-world datasets, namely Maven, C6, and C36. The results show the ability of the modified FS method to improve the performance of the proposed framework for ED tasks over other existing methods.