Exploring Multiscale Causal Interventions for Burned Area Estimation

Abstract. Understanding complex causal interactions between local, continental, and global drivers remains a significant challenge in wildfire prediction systems. This study implements a causal inference framework combining the Peter-Clark momentary conditional independence (PCMCI) algorithm with do-calculus interventions to analyse land-atmosphere feedback mechanisms influencing wildfire dynamics across South Asia (India, Pakistan, Myanmar, and adjacent regions). Time-series causal graphs derived from satellite and reanalysis data identified 500-hPa geopotential height anomalies (ΔZ₅₀₀) as the primary driver of surface aridity and wildfire incidence. Extreme scenario simulations via do-operator perturbations revealed that artificially enhancing ΔZ₅₀₀ to the 100th percentile, representing intensified upper-tropospheric ridging, produced the most severe mean burned area outcomes. Under these conditions, mean extreme burned area reached ~4.2 log ha (~15,000 ha), exceeding impacts from other perturbed variables. The integration of PCMCI-derived causal networks with counterfactual analysis provides a novel methodology for disentangling multiscale wildfire drivers, offering critical insights into future climate-driven fire risks through explicit representation of teleconnection mechanisms.