Wildfire

Wildfires are unplanned and uncontrolled combustion of vegetated areas which can damage property and human life, and may have negative effects on the ecosystems where they occur. Predicting wildfires is an important part of mitigating their negative impact.

Anthropogenic–induced changes to temperature and precipitation patterns, including more prolonged periods of hotter, drier conditions and lower winter precipitation in some regions, increase the risk of wildfires in these locations. Such changes in ambient temperature directly perturb several cryological and hydrological environmental properties, for example, losses in snow depth/coverage due to reductions in precipitation and earlier snow melting. In Europe, hydrological models have predicted rates of evapotranspiration are likely to increase in-tact with temperature rises, which may have the indirect effect of reducing soil and vegetation moisture.

This project focused on the wetlands of Polesia, a floodplain habitat of 186,000km² comprised of peatland, wetlands, and forests which is especially susceptible to increasing wildfires. Polesia’s susceptibility to wildfire may lie in its peat bogs’ sensitivity to reductions in moisture, with drier land more likely to suffer from wildfire. Additionally, reductions in vegetation moisture increase the likelihood of combustion and hence wildfire. Furthermore, direct anthropogenic stresses including land drainage, purposeful fires, and logging have tangible effects upon wildfire frequency and ferocity.

Having recognised direct and indirect effects and their potential to exacerbate wildfires in the Polesian region, this project aimed to identify spatial relationships between wildfire distribution and cryological, climatic, and pedological drivers of wildfire. Therewith allowing wildfire prediction which would facilitate the mitigation of destructive wildfire occurrences. This was to be achieved in two steps; firstly, identification of target and predictor geospatial environmental datasets- one predictor, land-cover type, was derived from previous research commissioned by the British Trust for Ornithology (BTO), where a land-cover classification algorithm was specially trained on a sub-region of Polesia. Secondly, a convolutional neural network (CNN) was used to reduce the dimensionality of the datasets sampled from- the architecture was based on a U-Net originally described for biomedical image segmentation. The CNN was trained to successfully predict wildfire location on a small subset of imagery. Further work could investigate how increasing the size of the training dataset makes the tool more robust at predicting wildfire locations in a diverse range of settings.

All code can be found in our GitHub repository: https://github.com/ai4er-cdt/WildfireDistribution