Twitter has evolved into a pivotal platform for information exchange, particularly during emergencies. However, amidst the vast array of data, identifying tweets relevant to damage assessment remains a significant challenge. In response to this challenge, this study presents a novel approach designed to identify tweets related to damage assessment in times of crises. The challenge lies in sifting through an immense volume of data to isolate tweets pertinent to the specific event. Recent studies suggest that employing contextual word embedding approaches, such as transformers, rather than traditional context-free methods, can enhance the accuracy of disaster detection models. This study leverages multiple granularity level context representation at the character and word levels to bolster the efficiency of deep neural network techniques in distinguishing between disaster-related tweets and unrelated ones. Specifically, the weighted character representation, generated with the self-attention layer, is utilized to discern important information at the fine character level. Concurrently, Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) algorithms are employed in the word-level embedding to capture global context representation. The effectiveness of the proposed learning model is assessed by comparing it with existing models, utilizing evaluation measures viz., accuracy, F1 score, precision, and recall. The results demonstrate the effectiveness of our model compared to existing methods.