Abstarct: This research addresses the problem of automatically identifying disaster-related posts in social media data streams on X/Twitter, where the high volume of data, noise, linguistic ambiguity, informal language, short message length, and the heterogeneity of multimedia content make it difficult to detect relevant and actionable information. The main objective of this study is to design and evaluate a computational frxamework for the reliable and efficient detection of disaster-related posts, so that it can be used in time-sensitive conditions for event monitoring and supporting situational awareness. To address this problem, a set of classical and deep learning methods were investigated and evaluated, including TF-IDF-baxsed textual representations, traditional classifiers, transformer-baxsed text encoders such as BERT and DistilBERT, convolutional neural network-baxsed image encoders such as MobileNetV2 and EfficientNet, and multimodal fusion strategies for combining text and image information. The models were evaluated using metrics such as accuracy, precision, recall, and F1-score, across several experimental scenarios. The results of the study showed that the proposed frxamework achieved highly successful performance in detecting disaster-related posts and attained high accuracy together with a rapid convergence trend under different scenarios. The findings also indicate that the use of advanced textual and multimodal approaches can improve the reliability of disaster monitoring systems and reduce the burden of manual information processing in critical situations.