Social media platforms such as X (formerly Twitter) have become critical real-time information sources during natural disasters. The large volume of crisis-related posts makes manual analysis infeasible, and existing classification approaches often lack the granularity, interpretability, and spatial awareness required for effective humanitarian response. This paper presents an end-to-end disaster tweet classification system that categorizes tweets into five humanitarian categories using a fine-tuned all MiniLM-L6-v2 encoder with a lightweight classification head, inspired by the DeLTran15 framework. The system integrates gradient-based Explainable AI for token-level interpretation, a hybrid NLP pipeline for extracting actionable intelligence such as locations, casualty counts, and resource needs, a confidence based Human-in-the-Loop mechanism for low-confidence predictions, and a Geospatial Temporal Aggregation module that clusters tweets by location, computes consensus distributions, and generates situation reports with severity scoring using a 10-case decision matrix. Experimental results on the CrisisMMD dataset demonstrate an accuracy of 87.0% and a macro F1-score of 87.26%.

Disaster Tweet Classification,Explainable Artificial Intelligence (XAI),Natural Language Processing (NLP),Transformer Models,Geospatial-Temporal Aggregation,Human-in-the-Loop Learning,Crisis Informatics,Disaster Response Intelligence