Multiparameter Causal Models for the Estimation and Explainability of Wildfire Burned Area

Understanding the causal relationships among local, continental, and global variables is crucial for improving the accuracy and reliability of wildfire estimation. However, uncovering these complex interactions remains a challenge in wildfire modeling. This study employs a data-driven causal discovery approach using the Peter&#x2013;Clark algorithm with momentary conditional independence (PCMCI) to analyze the time-series data and construct causal graphs. The aim is to enhance understanding of land&#x2013;atmosphere interactions with wildfires in South Asia, including India, Pakistan, Myanmar, and surrounding regions. In addition, a do-calculus approach is applied to perturb multiscale input variables, estimating the burned area under worst-case conditions. Results highlight the dominant role of 500-hPa geopotential height anomalies (&#x0394;<italic>Z</italic>₅₀₀) in driving surface dryness and wildfire occurrence. By simulating the physical relationships between local- and global-scale variables, including oceanic and climatic indices with varying time lags, the model identifies &#x0394;<italic>Z</italic>₅₀₀ as the primary causal factor leading to increased burned area magnitude and frequency. When &#x0394;<italic>Z</italic>₅₀₀ is bootstrapped to the 100th percentile, indicating the development of a strong upper-tropospheric ridge, the estimated burned area rises significantly. Under this extreme scenario, the mean burned area reaches &#x223C;4.2 log ha (&#x223C;15 000 ha), while extreme cases exceed 5 log ha (&#x223C;100 000 ha), with a pairwise <italic>p</italic>-value &#x2264; 0.0001. This burned area surpasses that of all other input variables under equivalent worst-case conditions. By integrating PCMCI and do-calculus, this approach advances causal inference in wildfire modeling, providing valuable insights into the interactions governing wildfire dynamics and their potential future risks.