Harvesting Natural Disaster Reports from Social Media with 1D Convolutional Neural Network and Long Short-Term Memory

The employment of social media platforms is progressively assuming pivotal roles in natural disaster management as early warning and monitoring systems. In emergencies, social media users can post real-time information regarding the disaster they are experiencing, along with their location. In contrast, other users can access this information, and volunteers can search for these posts to render assistance. Nonetheless, the effective utilization of social media for monitoring natural disasters remains challenging due to the lack of automated and prompt tools for identifying natural disaster reports on social media. The analysis of social media data, particularly Twitter data, heavily relies on Natural Language Processing (NLP) algorithms to classify natural disaster reports obtained from eyewitnesses. Previous studies developed a classification model utilizing a 1D Convolutional Neural Network (CNN) with feature extraction based on three word embedding techniques: Word2Vec, fastText, and Glove. Although CNN can optimize features used in the classification process, it lacks the capacity to comprehend the structure of data sequences or sequential data. Thus, this research aims to develop a combined classification model of 1D CNN and Long Short-Term Memory (LSTM). LSTM can comprehend temporal features, such as the word order in a given document. The 1D CNN + LSTM model developed in this research exhibits performance of 82.83%, 88.33%, and 81.77% for Floods, Forest Fires, and Earthquakes. The paired t-test conducted shows a significant increase with a p-value of 0.0035. This research significantly contributes to identifying the 1D CNN + LSTM model that demonstrates the best performance in classifying natural disaster reports. By exploring the effectiveness of 1D CNN + LSTM models with different word embedding techniques and levels of contamination, this research offers various insights into the optimal method for harvesting natural disaster reports from social media. Moreover, this study establishes the fundamental groundwork for selecting an appropriate number of words for word padding techniques to improve classification performance.