Burned area, severity, ET, and other geospatial datasets
We obtained high resolution datasets on fire activities and landcover for the Western US (WUS). Burn area boundaries, extents, and burn severity were obtained from the Monitoring Trends in Burn Severity (MTBS) dataset which is mapped at ~30 m and contains large historical record of CONUS wildland fires during 1984-2021 (at the time of writing). We chose three burn severity classes of low, moderate, and high severity and each fire event was characterized by one of the classes based on the mode of the severity of all burned pixels. A total of 1732 fire events (burn area > 1 km2) across WUS were selected from this dataset that occurred between 2014 and 2020, representing multiple vegetation types and climates. Only the fires ignited from June to December of every year (natural/prescribed) were selected to ensure the immediate impacts coincide with respective water years. We further filtered out cases with large pre-fire biases in ET and precipitation between control and burn polygons (equation 1). This provided with a total of 1514 fires across WUS.
Estimates of ET were obtained from the OpenET multi-model dataset27 which provides a consistent and operational framework to deliver daily ET estimates at 30 m. OpenET is currently limited to the WUS, with a period of record from 2016 to present. This paper focused on the ALEXI/DisALEXI model products, a multi-scale surface energy balance modeling system designed to generate self-consistent flux assessments from field to regional/continental scales37. Estimates from three other OpenET models, namely GEESEBAL, eeMETRIC, and PT-JPL, and from MODIS 8-day global ET product at 500 m (MOD16A2)34 were also used. OpenET DisALEXI (Disaggregation of the Atmosphere-Land Exchange Inverse) uses Landsat land surface temperature (LST) to spatially disaggregate regional ET maps that themselves are based on coarser resolution ALEXI ET estimates38,39. Regional ALEXI ET data are produced at 4-km resolution over the U.S. using thermal imagery from the Geostationary Operational Environmental Satellites (GOES). LST is the primary remote sensing input to both ALEXI and DisALEXI, directly constraining estimates of sensible heat and net radiation, and indirectly latent heat by residual. Other satellite inputs are leaf area index (LAI) to govern the soil/canopy partitioning and albedo to compute net radiation. While DisALEXI calculates the two-source energy balance (TSEB) implemented at the time of morning Landsat overpass, the disaggregation is performed in terms of daily fluxes by iteratively adjusting the upper boundary condition in air temperature until the reaggregated daily fluxes match with the ALEXI pixel value40.
Landcover maps were derived from MODIS annual 500 m maps (MCD12Q1.061) using International Geosphere-Biosphere Programme (IGBP) classification for the year 2020. MODIS plant functional types were reclassified into forests (classes 1-6), savannas (classes 8-9), and grasslands (class 10) for the characterization of WUS fire events. MTBS burn severity and MODIS landcover maps were overlaid with burn area boundaries to identify the representative landcover and burn severity for each fire event based on the mode of all burned pixels. Ecoregions were identified using bounding extents from western U.S. National Ecological Observatory Network (NEON) domains. Topographic data was acquired using digital elevation model from Shuttle Radar Topography Mission (SRTM) at 30 m resolution.
Climatic dataset
We used standardized precipitation evapotranspiration index (SPEI) derived from the 4-km daily Gridded Surface Meteorological (GridMET) dataset to define droughts at a monthly timescale28. SPEI calculates the difference between precipitation and reference ET at various time scales (ranging from months to years) and denotes the magnitude of drought conditions standardized over long-term conditions of 1981-2016. We used 3-month SPEI, which integrates the climatic water balance (difference of precipitation and reference ET) over previous 90 days, to characterize pre- and post-fire water years with values less than –0.49 indicating anywhere between mild to extreme drought conditions.
Control polygon delineation
To accurately isolate the impact of fires on water and carbon balance for each fire event, we identified a control polygon that is free of disturbance from burns. Using MTBS burn extent polygon, we first define a buffer region that circumscribes the control polygon at a distance of 100 m outside burn boundary and proportional to the square root of burn area. A buffer surrounding burn polygon from all directions minimizes the differences in ET retrievals due to differences in terrain slopes and aspects. Next, a set of masks were applied to ensure that the retrievals in control polygon do not get biased due to factors other than the fire disturbance. The masks include neighboring burn scars, water, and representative landcovers other than the major landcover within the burned area. Control polygon pixels with elevation differences larger than 200 m from burn polygon were also masked. The steps are summarized in Figure S4. The calculations were carried out using cloud computing platform of Google Earth Engine.
Calculation of impacts on water balance
We calculated estimates of water balance using precipitation from the reanalysis product of PRISM (Parameter-elevation Relationships on Independent Slopes Model)41 and gridded ET from the four OpenET models27 and the MODIS product34. Next, we quantify the shift in water balance due to fire disturbance as the difference in water balance over fire impacted region relative to the control polygon. To account for biases in ET or precipitation retrievals due to imperfect delineation of control polygon, we subtract the water balance difference between burn and control polygon calculated over the pre-fire year. Thus, water balance shift in post-fire year (in units of depth) is formulated as,
The estimates of volume were then aggregated over individual ecoregions to examine the sensitivity of vegetation disturbance to landcover, burn severity, and standardized drought index. Changes to volume were reported as percent recovery across subsequent water years (calculated as relative difference between the volume in two years).
The major uncertainties in our results stem from (i) bias introduced in calculating suppression when control polygons may include impacts of fire or non-representative landcover, terrain or other factors that affect ET or GPP retrievals, (ii) the mismatch in spatial resolutions of ET, precipitation, and SPEI, and (ii) inherent uncertainties in remote sensing observations of ET, GPP, and precipitation.
Trend analysis for drought impacts on ET
To quantify sensitivity of recovery with drought severity over three post-fire historical seasons, we performed trend analysis between NWSI and annual SPEI-90 for 761 fire events during 2016 to 2018. Characterized by burn severity, the trend was obtained between suppression over the three post-fire years and the respective SPEI across each landcover type. The slope of linear regression was obtained for fires in each landcover type and burn severity. Statistical significance of the slope was tested using two-tailed Wald Test with t-distribution of the test statistic.
Scenario analysis
To decouple the response of each landcover type to disturbances from fires and droughts, we identified two scenarios based on prevailing and following drought conditions for each major landcover type (forests/savannas/grasslands). Only moderate and high severity fires were considered in the scenario analysis to ensure fire disturbance was significant enough to have the impacts persist over multiple seasons. First, a drought scenario (DRS) entailed fire events that are preceded by dry extremes in the pre-fire year and are also affected by drought in atleast two of the three post-fire years. The response to DRS was compared with a baseline scenario (BLS) where the pre-fire year was a normal (or wet) year while post-fire years suffered either none or mild drought conditions (mild drought year in BLS were only allowed for forests where number of fire events in years without drought were less than ten). A year was classified as affected by drought when the median and mean of SPEI is less than -0.49 (anywhere from mild to extreme droughts) and labelled as non-drought year otherwise. Post-fire recovery of water balance is thus quantified as percentage reduction in mean NWSI of fire events relative to the first post-fire year.