11:45 〜 12:00
[HTT15-11] Geospatial Big Data and AI for Smart Humanitarian Mapping during Natural Disasters
★Invited Papers
キーワード：GIS, Disaster Resilience, AI, Big Data, Cartography, Sustainability
Human communities worldwide have witnessed a series of natural disasters. Despite enormous efforts to predict disasters’ occurrences, intensities, and impacts, disaster damages have consistently increased over the past few decades, partially due to the uncertain and complex nature of human-disaster interactions. Building resilient communities that can withstand future disasters has become a global challenge and priority that requires innovative, effective solutions.
A crucial practice in building disaster resilience is humanitarian mapping to identify affected areas, infrastructural and social impacts, and humanitarian needs during disasters. Advances in geospatial big data and artificial intelligence (AI) have opened new frontiers in timely, accurate, and detailed humanitarian mapping. Geospatial big data, derived from sources such as satellite imagery, street-view images, social media, and crowdsourced platforms, offers unprecedented volumes of spatially and temporally rich information. When coupled with AI, these data can be transformed into actionable insights. The integration of these technologies into humanitarian mapping facilitates rapid damage assessments, the identification of victims and their needs, and prioritizing resources, paving the way for disaster resilience.
This research aims to foster Smart Humanitarian Mapping during different disasters using geospatial big data and newly developed AI algorithms. The objectives are three-fold. First, this work establishes a comprehensive framework to employ geospatial big data and AI to enhance humanitarian mapping efforts. Second, we showcase the use of geospatial big data and AI in humanitarian mapping through three case studies, including (1) using social media and large language models (LLM) to map rescue requests during hurricanes, (2) applying nighttime light (NTL) remote sensing and machine learning for power outage mapping during winter storms, and (3) employing bi-temporal street-view images and dual-channel large vision models for hyperlocal damage estimations during coastal storms. Finally, this investigation identifies the challenges associated with integrating geospatial big data and AI in humanitarian mapping research and practice and proposes possible solutions.
The first study uses social media and fine-tuned LLM to map rescue requests during Hurricane Irma in Florida, U.S. Two novel AI models, VictimFinder and TopoBERT, were developed to search and geo-locate victims requesting help on Twitter/X. VictimFinder and TopoBERT achieved F-1 scores of 0.919 and 0.854. The second study applies NTL and advanced image processing to map power outages during the 2021 Winter Storm Uri in Texas, U.S. NASA’s Black Marble daily NTL images were used. Statistical adjustments were applied to mitigate the effects of viewing angle and snow reflection. Power outage was detected by comparing storm-time and normal condition NTL images, achieving an R-squared value of 0.42. The third study collected 2249 pairs of street-view images captured before and after the 2024 Hurricane Milton in Horseshoe Beach, Florida, U.S., for experiments. Dual channel fine-tuned pre-trained image classification models based on Swin Transformer and ConvNeXt were proposed and evaluated. The results show that the Dual-ConvNeXt + Cross-Attention model outperforms other models with an accuracy of 76.91% in identifying and classifying street-level damages.
This research highlights the transformative potential of geospatial big data and AI in advancing smart humanitarian mapping. By integrating social media data, nighttime remote sensing, and street-view images with cutting-edge AI models, the work enhanced our capabilities in mapping rescue requests, power outages, and infrastructure damages during disasters. These efforts not only provide actionable insights for disaster response but also address challenges in implementing these technologies, contributing to the development of more resilient communities.
A crucial practice in building disaster resilience is humanitarian mapping to identify affected areas, infrastructural and social impacts, and humanitarian needs during disasters. Advances in geospatial big data and artificial intelligence (AI) have opened new frontiers in timely, accurate, and detailed humanitarian mapping. Geospatial big data, derived from sources such as satellite imagery, street-view images, social media, and crowdsourced platforms, offers unprecedented volumes of spatially and temporally rich information. When coupled with AI, these data can be transformed into actionable insights. The integration of these technologies into humanitarian mapping facilitates rapid damage assessments, the identification of victims and their needs, and prioritizing resources, paving the way for disaster resilience.
This research aims to foster Smart Humanitarian Mapping during different disasters using geospatial big data and newly developed AI algorithms. The objectives are three-fold. First, this work establishes a comprehensive framework to employ geospatial big data and AI to enhance humanitarian mapping efforts. Second, we showcase the use of geospatial big data and AI in humanitarian mapping through three case studies, including (1) using social media and large language models (LLM) to map rescue requests during hurricanes, (2) applying nighttime light (NTL) remote sensing and machine learning for power outage mapping during winter storms, and (3) employing bi-temporal street-view images and dual-channel large vision models for hyperlocal damage estimations during coastal storms. Finally, this investigation identifies the challenges associated with integrating geospatial big data and AI in humanitarian mapping research and practice and proposes possible solutions.
The first study uses social media and fine-tuned LLM to map rescue requests during Hurricane Irma in Florida, U.S. Two novel AI models, VictimFinder and TopoBERT, were developed to search and geo-locate victims requesting help on Twitter/X. VictimFinder and TopoBERT achieved F-1 scores of 0.919 and 0.854. The second study applies NTL and advanced image processing to map power outages during the 2021 Winter Storm Uri in Texas, U.S. NASA’s Black Marble daily NTL images were used. Statistical adjustments were applied to mitigate the effects of viewing angle and snow reflection. Power outage was detected by comparing storm-time and normal condition NTL images, achieving an R-squared value of 0.42. The third study collected 2249 pairs of street-view images captured before and after the 2024 Hurricane Milton in Horseshoe Beach, Florida, U.S., for experiments. Dual channel fine-tuned pre-trained image classification models based on Swin Transformer and ConvNeXt were proposed and evaluated. The results show that the Dual-ConvNeXt + Cross-Attention model outperforms other models with an accuracy of 76.91% in identifying and classifying street-level damages.
This research highlights the transformative potential of geospatial big data and AI in advancing smart humanitarian mapping. By integrating social media data, nighttime remote sensing, and street-view images with cutting-edge AI models, the work enhanced our capabilities in mapping rescue requests, power outages, and infrastructure damages during disasters. These efforts not only provide actionable insights for disaster response but also address challenges in implementing these technologies, contributing to the development of more resilient communities.