A multimodal system to support disaster risk management
Original (Portuguese): Sistema Multimodal de Apoio ao Gerenciamento de Riscos de Desastres
Paper · 2026 · Hosana Iasmin Castro dos Santos Lucena, Gabriel Rocha dos Santos, Jady Lima da Silva et al.
Summary
This demo paper presents NOAH, a multimodal AI system for disaster risk management in Brazilian cities. The authors found that the civil defense office in Natal-RN logs threats and incidents by hand, first on paper and then in Excel. Under that workflow, categorization depends on individual judgment and automated statistical or spatial analysis is hard. NOAH lets citizens report incidents in Portuguese through a WhatsApp bot using text, voice or photos:
- Audio is transcribed with Whisper Small, and the user confirms the transcript.
- Text is categorized by a BERTopic model adapted to disaster classes.
- Images are classified by a ResNet50 CNN.
The bot then asks for the user's location and saves a structured record that appears on a web dashboard for civil defense staff. A second bot option uses RAG over official contingency plans to give guidance. The paper reports no quantitative evaluation; the authors plan to gather and curate real incident data for proper validation. The work matters because annotated crisis resources such as HumAID are mostly English, leaving Brazilian Portuguese and its regional varieties underserved.
Key points
- Pipeline components: Whisper Small (speech-to-text), BERTopic (Grootendorst, 2022) for text categories, ResNet50 for scene types such as floods, fires and landslides, and RAG over official contingency plans.
- Interface: citizens use only WhatsApp; staff view geolocated structured reports on a web platform.
- Motivation: the UNDRR (2025) estimates that people born in 2025 have about an 86% chance of living through a catastrophic flood, compared with 63% for those born in 1990. Brazilian funding for municipal risk-reduction plans (PMRR) and climate adaptation plans, which began with Lei 12.608/2012, has increased in 2024–2026.
- Gap addressed: existing crisis datasets (e.g., HumAID from CrisisNLP) focus on English, and an earlier reporting app (Nik Nazli et al., 2016) had no AI processing.
- Limitation: no metrics are reported, and evaluation on curated real-world data is left for future work.
Fields, topics and concepts
📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works
🧭 Topics 1Research subjects and languages the work is about.Vision & language · 657 works
Cite
Cite as Hosana Iasmin Castro dos Santos Lucena, Gabriel Rocha dos Santos, Jady Lima da Silva, Ricardo José Matos de Carvalho and Patrick Terrematte. 2026. Sistema Multimodal de Apoio ao Gerenciamento de Riscos de Desastres. Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 2.
@inproceedings{lucena2026sistema,
title = {{Sistema Multimodal de Apoio ao Gerenciamento de Riscos de Desastres}},
author = {Hosana Iasmin Castro {dos Santos Lucena}, Gabriel Rocha {dos Santos}, Jady Lima {da Silva}, Ricardo José Matos {de Carvalho} and Patrick Terrematte},
booktitle = {Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 2},
year = {2026},
url = {https://aclanthology.org/2026.propor-2.7/}
}References
- Muhammad Imran Ferda Ofli Firoj Alam, Umair Qazi. 2021. Humaid: Human-annotated disaster incidents data from twitter. In 15th International Conference on Web and Social Media (ICWSM). find
- Matilde Martins Lencastre Godinho. 2024. The impact of natural language processing in disaster management: A systematic literature review. Master’s thesis, Universidade NOVA de Lisboa (Portugal). find
- Maarten Grootendorst. 2022. Bertopic: Neural topic modeling with a class-based tf-idf procedure. Preprint, arXiv:2203.05794.
Updated
This page in Markdown → https://md.research.elamur.ai/free/en/papers/lucena-2026-sistema-multimodal-de-apoio-ao/