How Loud Rumbles Hit Newsstands: A Data Analysis of Coverage and Spatial Bias in German News about Landslides Around the World

Problem Statement
Landslides are destructive, potentially fatal events that significantly impact socioeconomic stability. While global inventories like EM-DAT and the Global Landslide Catalog (GLC) exist, they suffer from spatial biases, limited temporal coverage, missing data, and strict inclusion criteria (e.g., minimum fatality thresholds) that lead to incompleteness. Conversely, news archives offer a rich, alternative source of disaster information but are subject to latent selection criteria, such as geographic proximity and narrative appeal, which can distort the representation of global risks. There is a need to systematically analyze how international disasters are reported in specific media ecosystems to understand these inequalities and potentially enrich existing disaster databases.

Methodology
The authors constructed a large-scale dataset of German news articles concerning landslides abroad between 2000 and 2024.

• Data Collection: A corpus of 317,036 German articles was retrieved from the wiso-net aggregator using 17 landslide-related keywords. After preprocessing, 88,940 candidate documents remained.
• Annotation and Validation: A random sample of 450 documents was manually annotated for relevance (actual landslide occurrence) and country of occurrence. The human-annotated sample served as a gold standard, with 58.6% of documents deemed relevant.
• Automated Processing: An open-weight, locally run Large Language Model (LLM), mistralai/devstral-small-2-2512, was employed to geolocate the remaining documents. The model extracted ISO-3 country codes for specific past or ongoing landslide events.
  • Performance: On the gold standard, the model achieved a binary precision of 0.814, recall of 0.947, and F1-score of 0.875 for relevance. Geolocation accuracy was 0.837 for exact matches and 0.924 for overlaps.
• Event Segmentation: The resulting time series of daily document counts for each country was segmented into "news events." A news event was defined as a contiguous cluster of active days, allowing a maximum of four sequential inactive days between active days.
• Comparative Analysis: The study compared the distribution of identified news events against two external measures:
  1. EM-DAT: The International Disaster Database (observed events).
  2. WB-GLHM: The World Bank's Global Landslide Hazard Map (predicted annual frequency).
• Salience and Divergence: The authors computed "salience scores" (log-transformed, min-max scaled) for news events and external measures. They calculated the residuals of a linear regression (news salience vs. external salience) to determine "divergence." Positive divergence indicates overreporting; negative divergence indicates underreporting relative to external baselines.

Key Results

• Dataset Scale: The final dataset comprises 59,504 documents from 240 outlets covering 178 countries, resulting in 5,479 active days and 5,543 identified news events.
• Concentration: Coverage is highly concentrated; the top 10 countries account for 56.4% of all instances, and the top 29 countries cover 82.3%. Italy, Switzerland, Austria, China, and the USA are the most frequently reported.
• Spatial Bias:
  • Overreporting: Countries in the Global North, particularly in Southern and Western Europe, and high-income countries generally receive more media attention than expected based on their susceptibility to landslides (as measured by EM-DAT and WB-GLHM).
  • Underreporting: Parts of Asia (specifically South-Eastern Asia) and middle-income countries receive less attention than their risk profiles suggest.
  • Specific Extremes: Spain, Switzerland, Austria, Poland, and the UK show the highest positive divergence (overreporting) against EM-DAT. Conversely, Côte d'Ivoire, North Macedonia, Congo, Tajikistan, and Azerbaijan show the most extreme negative divergence (underreporting).
• Temporal Distribution: While the total number of documents shows a temporal bias toward recent years, the distribution of identified news events does not show a clear bias toward recent years, suggesting the segmentation method effectively normalized the data.
• Continental Disparities: Europe is significantly overrepresented in news events compared to external sources, whereas Asia is underrepresented.

Significance and Claims
The paper positions itself as a foundational step toward understanding how social inequalities are reflected and reinforced by disaster coverage in Global North media systems.

• Database Enrichment: The study demonstrates that news archives can serve as a valuable source to enrich incomplete disaster databases, providing a "text-as-data" paradigm for climate impact and adaptation research.
• Revealing Inequalities: By quantifying the divergence between media attention and actual/predicted hazard occurrence, the authors reveal systematic biases. They argue that German-speaking audiences receive a "partially distorted picture of global landslide risk," which may limit public understanding of the true geographic scale of disasters.
• Implications: The authors suggest that such skewed coverage could reduce political will and public support for international climate adaptation financing and risk reduction initiatives.
• Future Directions: The paper modestly outlines future work, including modeling coverage based on socio-economic and cultural proximity, distinguishing between past/ongoing/future event types, and expanding the analysis to other hazards like wildfires and floods. It explicitly notes limitations, such as the LLM's lack of generalization to other hazards and the reliance on country-level rather than regional spatial analysis.