Wildfire likelihood in Canadian treed peatlands based on remote-sensing time-series of surface conditions

Abstract

While peatland fires are less common than upland fires, long periods of moisture deficit can increase fuel availability and lead to greater fire risks in peatlands. Peatland fire likelihood is crucial as these ecosystems serve as natural fire breaks under normal hydrological conditions but can easily turn into fuel sources under drought conditions. Furthermore, peatland fires have significant implications for atmospheric pollution, releasing fine particulates, carbon, and mercury emissions. Current frameworks for estimating wildfire likelihood predominantly rely on fuel-moisture indices developed for upland environments, which cannot adequately capture the hydrological and fuel-moisture conditions specific to peatlands. Additionally, common fuel-moisture indices are reliant on station-derived meteorological data, which are scarce in remote areas and therefore imprecise over large portions of sub-Arctic Canada and the boreal forest.Remote sensing data represents a proxy for near real-time surface conditions for remote locations, such as moisture and vegetation, where a time series of remotely-sensed metrics provides information about patterns and trends in local conditions over longer periods. With this in mind, we used SAR (Sentinel-1), multispectral (Sentinel-2) and human infrastructure data as a predictor of wildfire probability in treed peatlands of Canada. We developed a fire probability model based on metrics drawn from a seasonally decomposed time series of SAR backscatter and common multispectral indices of greenness, soil moisture and vegetation moisture. Sample sites were derived from an existing treed peatland map, then the Canadian Forest Service National Burned Area Composite dataset was used to identify fire events within those areas. This study presents a robust and versatile framework for training fire probability models that leverage remote sensing data and temporally-invariant ancillary predictors. Fire probability in peatland locations, given current conditions, was predicted through a model algorithm called extreme gradient boosting. To obtain fire probabilities over a four-week timeframe, the output of the binary classifier was calibrated, transforming its raw scores into more accurate probabilities. This approach showcases the effectiveness of machine learning methods in harnessing seasonally-decomposed remotely sensed time-series to improve fire prediction.The most important predictors of fire likelihood in treed peatlands were found to be multispectral moisture-related indices (Tasseled cap Wetness and Normalized Difference Moisture Index) and the distance to human infrastructure, which may indicate an overrepresentation of human-caused fires in the predictions. Using subsets of predictors, different iterations of the model demonstrate that incorporating SAR backscatter data can enhance the accuracy of fire likelihood predictions. However, models relying solely on SAR data exhibited lower performance, indicating limited applicability of SAR backscatter in treed peatlands without more advanced filtering and a correction method.The fire likelihood model presented here adds seasonally-detrended time series metrics from remote sensing observations to inform fire risk, independent from meteorological data. Such models could be implemented for a near-real time assessment of fire conditions that could complement the existing framework for fire risk predictions in Canada.