<p>The municipalities of Qeqertalik and Qeqqata in western Greenland experienced two wildfires in July 2017 and July 2019, respectively. Both fires occurred near Sisimiut, the second largest city in Greenland, with the ignition site of the July 2019 wildfire along the Arctic Circle Trail. These Arctic fires vary in fuels and burning behaviour from other high northern latitude fires due to unique flora, specifically the lack of extensive grasses, shrubbery, and more vascular vegetation, and presence of deep vertical beds of carbon-rich humus. The purpose of this research was to create wildfire risk models scalable across the Arctic landscapes of Greenland. We test multiple wildfire risk models based on expert-derived weighted matrix and four geostatistical techniques: Equal Influence (eq_infl), Multiple Logistic Regression (MLR), Geographically Weighted Regression (GWR) and Generalized Geographically Weighted Regression.The eq_infl model applied an even influence of each landscape characteristics. Two MLR models were developed, one using all the available data for the peninsula where the wildfire occurred (MLR_full) and the other which used an equal randomly chosen 50,000 pixel subset of both the burned area and unburned area (MLR_sub) immediately surrounding the 2017 Qeqertalik wildfire.The optimum model was selected in a stepwise fashion for both MLR models using AIC. GWR and GGWR models were derived from the MLR_sub, to avoid multicollinearity. Landscape characteristics for the wildfire risk models relied on open source remotely sensed data like ~20 m synthetic aperture radar imagery from the European Space Agency Sentinel-1 for soil moisture; elevation, slope, and aspect derived from the 10 m Arctic DEM provided by the U.S. National Geospatial Intelligence Agency (NGA) and National Science Foundation (NSF); vegetation fuel beds from the Global Fuelbed Dataset; normalized difference vegetation indices (NDVI) from 20 m Sentinel-2 served as proxies for vegetation condition; and soil carbon information from the 250 m SoilsGrid product was used to indicate likelihood of humus combustion. The nominal spatial resolution of each wildfire risk model was 20 m, after resampling of data. The optimum wildfire risk model was the model that displayed the greatest fire risk within the 2017 burned area. The average fire risks for each model were compared for significant difference in the mean fire risk using an ANOVA and Tukey's Post hoc. Average predicted fire risks by our models were compared to 2017 and 2019 burned areas visually digitized from 10 m Sentinel-2 data. The MLR_full model best represented the burned area of the 2017 Qeqertalik wildfire, though with an R<sup>2</sup> of 0.232, this leaves large amounts of variation unexplained. This is not surprising as wildfires in Greenland are uncommon and applying traditional fire risk approaches may not accurately represent the real-world. We can interpret from the results of the MLR_full model that landscapes across western Greenland have the potential to burn in a similar manner to the 2017 and 2019 wildfires.</p>
Deep Learning for Monitoring Agricultural Drought in South Asia Using Remote Sensing Data
