Patterns of mega-forest fires in east Siberia will become less predictable with climate warming
Very large fires covering tens to hundreds of hectares, termed mega-fires, have become a prominent feature of fire regime in taiga forests worldwide, and in Siberia in particular. Here, we applied an array of machine learning algorithms and statistical methods to estimate the relative importance of...
| Published in: | Environmental Advances |
|---|---|
| Main Authors: | , , , , , |
| Format: | Article in Journal/Newspaper |
| Language: | English |
| Published: |
Elsevier
2021
|
| Subjects: | |
| Online Access: | https://doi.org/10.1016/j.envadv.2021.100041 https://doaj.org/article/c33a494a91904ff39204e40542c00138 |
| Summary: | Very large fires covering tens to hundreds of hectares, termed mega-fires, have become a prominent feature of fire regime in taiga forests worldwide, and in Siberia in particular. Here, we applied an array of machine learning algorithms and statistical methods to estimate the relative importance of various factors in observed patterns of Eastern Siberian fires mapped with satellite data. More specifically, we tested linkages of “hot spot” ignitions with 42 variables representing landscape characteristics, climatic, and anthropogenic factors, such as human population density, locations of settlements and road networks. Analysis of data spanning seventeen years (2001–2017) showed that during low or moderately high fire seasons, models with full set of variables predict locations of fires with a very high probability (AUC = 95%). Sensitivity, or the ratio of correctly predicted fire pixels to the total number of pixels analyzed, declined to 30–40% during warm and dry years of increased fire activity, especially in models driven by anthropogenic variables only. This analysis demonstrates that if warming in Eastern Siberia continues, forest fires will become not only more frequent but also less predictable. We explain this by examining model performance as a function of either temperature or precipitation. This effect from climate makes it nearly impossible to segregate ignition points from locations, which were burnt several hours or even several days earlier. An increase in secondary burnt locations makes it difficult for machine learning algorithms to establish causality links with anthropogenic and other groups of variables. |
|---|