Enhanced Disaster Management by Leveraging Deep Learning Techniques

In recent years, the proliferation of social media platforms has created a valuable source of real-time data for disaster management. Project aims to improve situational awareness and response efforts through the efficient classification of social media-related message and image content during disasters. Well known Machine learning algorithms like Decision Trees are used for text classification but these methods have their own limitations which include the inability to deal with complex data structures and enforce a smooth scaling to larger datasets. To overcome these challenges, proposed an enhanced model leveraging deep learning techniques for image classification, specifically utilizing the more advanced ResNet50 architecture. Further, as an extension of the uni-modal classification, a multi-modal no-attention model has been employed for classifying disaster related post images. The Decision Tree performed well demonstrating an accuracy of 62%, with different precision, recall and F1-scores across different categories of disaster in its classification report. Image classification performed using ResNet50 got 63.74% accuracy and F1-score of 54.55%. To improve the classifications, a multi-modal classification model which concatenates Bi-Directional Long Short-Term Memory (Bi-LSTM) for text and Visual Geometry Group (VGG16) for images is used. This multi-model allows different data types to be processed together, leading to more comprehensive models of disaster classification.