Bayesian Network Analysis of Land-Atmosphere Interactions Affecting Burned Areas in India During the 2022 South Asia Heatwave

This study addresses discerning causal relationships in complex systems, a key aspect of interpretable machine learning. It focuses on the unusual and intense early summer weather in South Asia during April and May 2022 that led to an increased number of forest fires. This work employs a Bayesian network (BN), constructed using the NOTEARS algorithm, to analyse the contribution of various land and atmospheric variables on the extent of burned areas. In a scenario analysis using peak values of 300-hPa meridional circulation index and 500-hPa Geopotential Height Anomalies, indicative of a strong atmospheric block, the likelihood of large burned areas (>3.06 log ha or >1150 ha) increases from 36.6% to 41.6%. This is due to a rise of conditional probabilities in the Vapor Pressure Deficit (VPD) (> 5.21 kPa) by 24.8%, and the Land Surface Temperature (LST) (>45.7°C) by 15.6%. In addition, sensitivity and spatial analyses indicate that extreme dry conditions, characterized by high LST and VPD due to the trapping effects of the omega block jet stream pattern, were the primary factors influencing the extent of burned areas during the 2022 South Asia heatwave.