Towards better social crisis data with HERMES: Hybrid sensing for EmeRgency ManagEment System

TL;DR

This work tackles the challenge of enriching social crisis data by combining opportunistic OSN data with targeted participatory elicitation in a hybrid sensing framework. The HERMES system automates earthquake-triggered data collection, applying RCNN-based filtering, witness detection via SVM, geoparsing with Geo-Semantic-Parsing, and bot-enabled targeted questioning to gather richer, more actionable information in real time. Real-world experiments over 436 earthquakes show substantial gains in data volume and quality, including up to +20% more data, up to 7x higher geographic density, up to 18x greater geographic variety, and up to 30% improved coverage, with faster response times (Δt ≤ 28 minutes). HERMES also demonstrates higher witness engagement than USGS-DYFI in some non-US events, underscoring the potential for integrating hybrid sensing into emergency management workflows while addressing GDPR and ethical considerations.

Abstract

People involved in mass emergencies increasingly publish information-rich contents in online social networks (OSNs), thus acting as a distributed and resilient network of human sensors. In this work we present HERMES, a system designed to enrich the information spontaneously disclosed by OSN users in the aftermath of disasters. HERMES leverages a mixed data collection strategy, called hybrid sensing, and state-of-the-art AI techniques. Evaluated in real-world emergencies, HERMES proved to increase: (i) the amount of the available damage information; (ii) the density (up to 7x) and the variety (up to 18x) of the retrieved geographic information; (iii) the geographic coverage (up to 30%) and granularity.

Figures (10)

• Figure 1: System architecture of HERMES, a Hybrid sensing for EmeRgency ManagEment System.
• Figure 2: UML sequence diagram of the HERMES system. $\Delta_t$ represents the latency between the sending of our targeted questions and the possible user replies.
• Figure 3: Architecture of the Recurrent Convolutional Neural Networks used to train the models for message filtering and damage assessment steps.
• Figure 4: Examples of "conversations" with Twitter users. Conversations are composed of 3 tweets: (i) the topmost tweet is a spontaneous message collected with opportunistic sensing; (ii) the middle tweet is the targeted question automatically sent by our system, and (iii) the bottom tweet is the user reply to our question.
• Figure 5: Ratio of tweets reporting the presence of damage in solicited replies (reply2damage) with respect to spontaneous relevant tweets. Statistical significance of comparisons is evaluated by means of T-tests, with ***: $p < 0.01$; **: $p < 0.05$; *: $p < 0.1$.