Utilization of social media in floods assessment using data mining techniques

Area of Science:

• Environmental Science
• Data Science
• Disaster Management

Background:

• Floods cause significant human and financial losses globally.
• Arid regions face challenges in flood monitoring due to the unreliability of traditional flow gauges.
• Social media platforms offer a potential source of real-time data for disaster assessment.

Purpose of the Study:

• To investigate the reliability and quality of flood-related data from social media in arid regions.
• To evaluate the effectiveness of Machine Learning (ML) approaches in analyzing social media data for flood events.
• To establish social media data as a viable alternative to scarce flow gauge data.

Main Methods:

• Collected and analyzed social media data (text, images, videos) from a 2016 UAE flash flood event.
• Utilized ResNet50 model with VGG-16 architecture to convert digital data into numerical values.
• Employed various ML algorithms for data classification and evaluated performance using ROC curves and AUC methods.
• Validated social media data against precipitation/rainfall data.

Main Results:

• Random Forest classifier achieved 80.18% accuracy for image and video analysis.
• A significant correlation was found between rainfall and the volume of social media posts.
• YouTube videos demonstrated the highest data quality accuracy, followed by Facebook, Flickr, Twitter, and Instagram.
• ML models showed strong validity and accuracy through AUC, precision-recall curves, RMSE, and kappa statistics.

Conclusions:

• Social media data provides a valuable and reliable alternative for flood monitoring, especially in data-scarce arid regions.
• Machine learning techniques are effective in processing and analyzing diverse social media data for disaster management.
• The study confirms the potential of leveraging user-generated content for enhanced flood event understanding and response.