Rapid Disaster Damage Estimation Based On Geo-referenced Social Media Data

In the aftermath of disasters, the spatial distribution of economic damage is critical to timely and efficient response and assistance delivery. However, the current estimation methods heavily rely on manual efforts, e.g., on-site survey, which is a time-consuming and costly process. More importantly, the disadvantaged communities often experience more difficulties in recovery than its counterparts because people of low socioeconomic status face many barriers to access needed resources. Social media platforms allow citizens to express, communicate, and report disaster impact in the formats of geo-referenced text, image, and video in real time and from the bottom up. This paper proposes a deep neural network (DNN) based approach to estimate the economic damage based on geo-referenced social media content. This relationship can be expressed as function DNN(Social Media)=Damage. The training of such DNN models requires paired numerical social media content and damage data from the same location. In this way, the fully trained models can be deployed to estimate the damage across regions.

Using the Texas winter storm Uri that occurred in February, 2021 as an example, a DNN is designed to take in the vectorized text data and output the damage in Dollars. For the data collection, the text data is extracted from the Twitter database using a list of keywords and the damage data is retrieved from the OpenFEMA database. Considering the FEMA damage data is published on the zip code level, this DNN is customized to predict the average damage at the same spatial granularity. Altogether, 338 paired data points, i.e., zip codes marked with grey, are used to train, validate, and test the model. Figure 1 visualizes the error percentage (intensity color coded) calculated by dividing the estimation discrepancy over FEMA data. On average, this model achieves a 68.42% accuracy. The red regions (rural areas in west Texas) indicate zip codes that have no FEMA surveys (communities in disadvantage) but have social media data (hence the DNN model can be used to estimate the damage). The results reveal more damages on the Northwestern Texas than Central Texas. The areas with no geo-referenced social media nor FEMA data are marked with light blue color. The proposed approach is a prototype of predicting disaster damages based on instant social media content.