The increasing frequency and severity of wildfires globally pose significant threats to ecosystems and human societies. This has spurred a dramatic rise in wildfire research, as evidenced by the more than fourfold increase in publications over the last two decades. However, the sheer volume of research presents challenges for traditional literature review methods, which often struggle to synthesize large datasets effectively. This paper addresses the need for efficient large-scale literature analysis by leveraging the capabilities of large language models (LLMs). LLMs, like GPT-3.5-turbo, offer the potential to overcome limitations of traditional methods by processing vast amounts of textual data and extracting key information, including geographical location and thematic classifications. The study aims to analyze the distribution of wildfire research across geographical regions and thematic areas to identify research gaps and inform future research priorities. The key questions addressed are: (1) What are the prevailing trends and preferences in wildfire research? (2) How do spatio-temporal variations in research paradigms manifest? (3) What are the implications of disparities in wildfire research for populations and socioeconomic development?

The authors acknowledge the limitations of traditional expert-based methods for analyzing the rapidly expanding body of wildfire research literature. They highlight the limitations in synthesizing the large volumes of data and identifying critical gaps in the research landscape. Previous studies examining specific aspects of wildfire research, like the use of remote sensing or the impact of climate change, have been cited in the paper. However, a comprehensive, large-scale analysis integrating geographical data and thematic categorization has been lacking. This study fills this gap by using an LLM approach to offer a broader perspective on current wildfire research.

The study employed GPT-3.5-turbo, an LLM, to analyze a dataset of over 60,000 peer-reviewed papers retrieved from the Web of Science database. The LLM was used to categorize publications into various themes related to wildfire research, including causes, consequences, and methodologies. A critical step in the methodology involved geoparsing: extracting and converting textual geographical information from titles and abstracts into numerical coordinates to map the geographical focus of publications. The authors describe a rigorous validation process to ensure the accuracy of the LLM's classification and geoparsing tasks, achieving an average F1 score of 0.85 through a cross-validation approach. Specific prompts were designed to extract information on major and minor disciplines, study area, study period, fire stage (pre-fire, actively burning, post-fire), and other key information. To ensure maximum coverage, prompts were repeatedly refined through testing. The resulting geoparsing yielded 60,488 articles for subsequent analysis. Supplementary data, such as burned area data from AVHRR-LTDR, fire emission data from GFED4.1s, global population data (GPWv4), and GDP data, were integrated to compare publication patterns with actual wildfire occurrences and socioeconomic indicators.

The analysis revealed significant geographical disparities in wildfire research. Western United States, despite representing less than 0.5% of global burned area, accounted for 15% of publications. Other regions with extensive wildfire activity, such as Siberia and Africa, were significantly underrepresented. Thematic analysis showed that "vegetation" was the most frequently discussed topic, with "forest fires" being the dominant sub-topic. Temporal analysis revealed shifts in research focus, with topics like "hydrological" and "atmospheric" impacts gaining prominence in recent decades, coinciding with advances in remote sensing technology. Analysis of author affiliations showed a strong bias towards high-income countries, indicating resource inequalities in funding for wildfire research and monitoring. A key finding is the considerable imbalance between research attention and actual wildfire activity. Regions with high levels of underrepresentation in wildfire research account for only 2% of the global GDP, reflecting limited economic capacity for both wildfire management and research. Biome-level analysis highlighted disparities in research attention towards grasslands and savannas in Africa and northern Australia, despite these biomes experiencing the highest percentage of global burned area (72%). The study also found imbalances in research focus on fire ignition sources, with human-caused fires increasingly dominating research in some regions, while natural causes remain understudied in others. The study quantifies imbalance levels across various factors, revealing a significant portion of the global population and socioeconomic development exposed to high imbalance levels (regions with significantly more burned area than research focus).

The findings reveal significant spatial and thematic biases in current wildfire research, potentially hindering effective wildfire management and mitigation strategies. The disproportionate focus on certain regions and themes may lead to a skewed understanding of global wildfire dynamics and their consequences. The underrepresentation of regions with extensive burned areas, particularly in developing countries, raises concerns about equity and resource allocation in wildfire research. The reliance on readily available data and research infrastructure, prevalent in wealthier countries, may bias the distribution of research towards these regions, regardless of their relative importance. This disparity in research can result in insufficient attention to regions most vulnerable to wildfire impacts and thus exacerbate existing inequalities. The study underscores the importance of promoting transdisciplinary collaborations and utilizing AI-aided approaches to overcome these biases and ensure a more comprehensive understanding of global wildfire phenomena.

This study demonstrates the efficacy of LLMs in analyzing large datasets of scientific literature to reveal previously hidden patterns. The findings highlight substantial disparities in the geographic focus and thematic scope of wildfire research, with underrepresentation of regions with significant wildfire activity and socio-economic vulnerability. The authors emphasize the importance of addressing these imbalances through increased funding, collaborative research efforts, and improved data sharing to achieve sustainable wildfire management. Future research could investigate the underlying causes of these biases and develop targeted strategies to promote more equitable and globally relevant wildfire research.

The study's reliance on the LLM's capabilities introduces potential limitations. The accuracy of the LLM's classifications and geoparsing depends on the quality of the input data and the effectiveness of the prompts used. While the authors conducted validation, residual errors are possible. The study also relies on existing datasets, which may have inherent limitations and inconsistencies in terms of spatial and temporal coverage and data quality. Finally, the study acknowledges the complexities of wildfire dynamics and the need for further investigation to better understand the diverse significance and impacts of wildfires across different regions and ecosystems.