Detecting Burned Area Anomalies with Isolation Forest in the Tropics: A Focus on Madagascar

Tropical forests are at high risk of dieback due to human-induced disturbances including forest fires, agricultural expansion, and logging. These disturbances can degrade the ecosystems, slow forest recovery, and disrupt the global carbon cycle, leading to irreversible changes or ‘tipping point’ in the Earth’s climate system – the point at which disruption to the climate potentially becomes irreversible. Early warning signals of tipping points for the Amazon rainforest and Greenland ice sheet have already been detected. Monitoring these forest ecosystems is crucial to mitigate future long-term consequences. In order to analyse the response of vegetation to disturbances, we must first identify such disturbances, ideally across the entire tropics over a long period of time. We must also carefully consider what we mean by a “disturbance” and it is not necessarily just the largest fire event. It may be that a significant disturbance is a modest fire event but in a region that does not typically experience burning or a fire event outside of the typical fire season. In both of those instances, we might expect the vegetation response to have different characteristics to those from regular, large burns.

In this study, we applied Isolation Forest (IF) algorithm to detect Burned Area (BA) anomaly and apply it to ESA FireCCI51 dataset (2001-2020) over IPCC AR6 defined land regions, with Madagascar as a case study region. IF identifies anomalies by considering how easily they can be isolated from the main distribution and allows us to introduce features beyond just the burned area itself (e.g., time and location of the fire). Explainable AI (SHAP) analysis was also performed to further understand the predicted BA anomaly. A higher number of BA anomalies were mostly linked to larger values of BA over the Tropics and in Madagascar, however, anomalies in BA are also affected by temporal and geographical factors other than the magnitude of BA. IF detected a high number of anomalies (>20) in the northern region of Madagascar which comparatively had lower BA values which could indicate deviation from seasonal fire patterns. These results were further explained by SHAP analysis which showed that BA was the main factor influencing prediction of BA anomaly but that time and location could play a significant role in some anomaly detections. This suggests that deviation from the typical fire seasonality was another factor contributing to anomaly detection. The high number of anomalies in these specific areas highlights the need for targeted fire management strategies so that policymakers can anticipate the long-term effects of climate change and human activity on tropical forests, guiding sustainable land use, conservation, and climate adaptation efforts in vulnerable regions.