3D word embedding vector feature extraction and hybrid CNN-LSTM for natural disaster reports identification

Abstract

Social media contain various information, such as natural disaster reports. Artificial intelligence is used to identify reports from eyewitnesses early for disaster warning systems. The artificial intelligence system includes a text classification model with feature extraction and classification algorithms. Word embedding-based feature extraction is widely used for 1-dimensional (1D) and 2-dimensional (2D) data, suitable for traditional or deep learning algorithms. However, applying feature extraction to 3-dimensional (3D) data for text classification is limited. Previous studies focused on word embedding for 1D, 2D, and 3D outputs with convolutional neural network (CNN). Yet, using 3D data and CNN did not perform well. Despite using CNN and 3D variants, identifying natural disaster reports remains below 80% accuracy. This research aims to improve identifying earthquakes, floods, and forest fires with 3D data and hybrid CNN long short-term memory (LSTM). The study found models with accuracies of 83.38%, 83.72%, and 89.03% for each disaster type. Hybrid CNN LSTM significantly enhanced identification compared to CNN alone, supported by statistical tests with P value less than 0.0001.