Wildfires in Greece : exploiting social media posts for fire detection, classification, and mapping through machine learning (ML) and natural language processing (NLP)
Loading...
Date
2026-06-01
Authors
Σαμαράς, Μάριος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Wildfires have become one of the most severe environmental threats in Greece, with increasing frequency and intensity driven by the ongoing climate crisis. In this context, leveraging social media posts to monitor fire events and to enhance situational awareness has emerged as a promising research approach. Transforming citizens into a decentralized network of digital “social sensors” provides a massive volume of valuable crowdsourced data that typically remains significantly underexplored. This diploma thesis proposes an end-to-end Machine Learning (ML) and Natural Language Processing (NLP) pipeline for the detection, classification and mapping of active fire events in Greece by exploiting Greek X (formerly Twitter) posts. To address the lack of publicly available Greek fire-related datasets, this study collected and manually annotated 3,277 X posts, while also constructing a separate ground truth dataset of 147 samples for geoparsing evaluation. The proposed workflow includes text preprocessing, binary classification to distinguish active fire-related posts, multilabel classification to identify specific fire types (wildland, urban or mixed) and a spatial analysis module for location extraction. Model evaluation emphasized on the F2-Score to minimize missed fire events, which is a vital requirement in emergency response. Four models were evaluated for the classification tasks, ranging from the traditional SVM/TF-IDF baselines to the Transformer-based models including the monolingual GreekBERT and the multilingual XLM-RoBERTa. In binary classification, the strongest performance on F2-Score was achieved by both GreekBERT (0.911±0.016) and XLM-RoBERTa (0.905±0.019), as a paired t-test showed no statistically significant differences between them. In multilabel classification, XLM-RoBERTa emerged with a Macro-F1 score of 0.874±0.024. For spatial analysis, this thesis presents a geoparsing module that combines Named Entity Recognition (NER), custom toponym disambiguation algorithms and Nominatim API for coordinate assignment. This optimized approach significantly improved operational geocoding accuracy from 54.68% to 80.58%. Finally, the DBSCAN algorithm supported the fire event clustering, and an interactive map of Greece and Cyprus was generated providing visual insights about detected fire incidents. Overall, this thesis bridges critical gaps in domestic crisis informatics by providing a scalable tool and by showcasing the successful exploitation of social media as a valuable complementary data source regarding fire events in Greece.
Description
Keywords
Crisis informatics, Fire detection, Social media, Natural language processing (NLP), Machine learning