The uncertainties and potential of locational-based social media for floodscape mapping

During a flood event, there has been an increasing number of observations recorded on Locational-based Social Media (LBSM) about the floodscape in near-real-time. Using the LBSM data harvested from X during Hurricane Harvey in 2017, this research adopted a computer vision (CV) approach that leveraged Visual Relationship Learning (VRL) by using a Faster Recurrent-Convolutional Neural Network (R-CNN) model to estimate Water Depth (WD) from a single flood scene. By validating against field surveying, the VRL yielded an accuracy of 66.8%. Compared to the stream gauge and High-Water Marks (HWM) data, the LBSM-WDs provide valuable information to supplement existing data sources in both space and time. The daily maximum WD was then interpolated to reconstruct the flood inundation maps during August 26–September 1, 2017. It was revealed that the median spatial error of LBSM posts could be more than half a kilometer. This research proposed a multi-object classifier that could yield flexible adaptability of various flood scenes with decent overall accuracy. This study also examined and mapped the significant differences of daily floodscape between LBSM and authoritative data. The findings provide insights to the uncertainties of LBSM and possible integration of multi-source geospatial data for future floodscape analysis.