Fire is a natural force of succession and evolution on western North American landscapes. Prior to European colonization, both individual species and ecosystems were maintained by episodic fire and evolved under the selective pressure of fire (Pausas and Keeley 2009; Stevens et al. 2020). However, the nature of fire is changing under anthropogenic climate change and historic fire suppression. Wildland fuels are more abundant and continuous today than they were prior to European settlement (Hagmann et al. 2021). At the same time, warmer and drier conditions are associated with more area burned(Westerling 2016; Abatzoglou et al. 2021) and higher severity fire (S A Parks and Abatzoglou 2020).
In light of shifting climatic conditions, there is a growing concern that some forests will not recover after fire. As more area burns at high severity (S A Parks and Abatzoglou 2020), stand replacing fire becomes more common on the landscape, which is the first step towards forest conversion (Sean A Parks et al. 2019). Failure to recruit a new cohort of trees in these areas ultimately leads to conversion (Davis et al. 2020). Small trees are more sensitive to climate extremes than their mature counterparts (Trouillier et al. 2019). Short term drought (Young et al. 2019), chronically dry conditions (Urza and Sibold 2017; Boucher et al. 2020; Stewart et al. 2021), and temperature extremes (Kemp et al. 2019) all limit seedling recruitment in early seral forests.
In the western United States, evidence is emerging of widespread conversion from forest to non-forest both currently and under future climate projections One third of forests in the Klamath region of northern California and southern Oregon could transition to shrublands by the end of the century (Serra-Diaz et al. 2018). In the Canadian boreal forests, wildfire is projected to facilitate the conversion of half of conifer forests to hardwoods or grass under some future climate scenarios (Stralberg et al. 2018). In the southwestern United States, increased fire activity is already associated with less conifer and greater shrub regeneration (Keyser et al. 2020).
Forests in the Northern Rocky Mountains may be more resistant to conversion than other systems in western North America. Under a 2-degree Celsius warming scenario, Montana forests are expected to experience less stand replacing fire and more recruitment after such fire than forests of the central and southern Rocky Mountains (Davis et al. 2020). Additionally, recent field surveys of burn scars in the Northern Rockies indicate abundant conifer regeneration and survival when close to a seed source (Clark-Wolf, Higuera, and Davis 2022). Similarly, Jaffe et al. (2023) documented abundant conifer regeneration across all fire histories in mixed conifer forests of the Selway-Bitterroot wilderness. Further west, Povak et al. (2020) also found significant regeneration following wildfire in mixed conifer forests of the Okanogan Highlands and Eastern Cascades, which they partially attribute to mild postfire weather conditions that may have reduced water stress on seedlings. Collectively, these findings indicate healthy forest recovery after wildfire in the region, at least in mixed-conifer forests. However, by mid-century, large swaths of the Northern Rockies are expected to experience summer heat extremes too warm to support widespread conifer seedling regeneration (Kemp et al. 2019). Northern Rockies forests currently sit at a crossroads: current conifer recruitment appears optimistic, but widespread conversion and community reorganization is possible before the end of the century.
Remote sensing is one of a few tools available to monitor forest recovery systematically across large land areas. In particular, Landsat data has been used extensively to study land surface change. Since the entire Landsat collection became free and publicly available in 2008, multiple metrics of analysis and methods for deriving ecological information from time series have been developed (Zhu 2017). Indices such as normalized difference vegetation index (NDVI) (S. Huang et al. 2021) and normalized burn ratio (NBR) (Bright, Hudak, Kennedy, Braaten, and Khalyani 2019) are spectrally derived metrics that are commonly proxies for vegetation through time. These metrics are simple, easy to calculate, and correlate with a wide range of vegetative characteristics in multiple biomes (S. Huang et al. 2021). However, the proliferation of cheap and accessible computing power and the development of new machine learning techniques have reduced reliance on spectral proxies and allowed for more direct modeling of forest attributes such as cover (Matasci, Hermosilla, Wulder, White, Coops, Hobart, and Zald 2018), height (Hudak et al. 2002; Pascual et al. 2010), and biomass (Sun et al. 2022; Pflugmacher et al. 2014) from satellite imagery.
Additionally, there has been extensive development in methodology for deriving disturbance information from a temporal series of metrics (Zhu 2017). LandtrendR (Kennedy, Yang, and Cohen 2010) uses a segmentation strategy to divide time series into components meant to reflect periods of landscape stasis and change and has been used extensively for diverse monitoring tasks since its development (Bright, Hudak, Kennedy, Braaten, and Henareh Khalyani 2019; de Jong et al. 2021; Runge, Nitze, and Grosse 2022). It has been used to detect disturbance when the locations are unknown (Cohen et al. 2018) and has been designed for sensitivity to both abrupt and subtle change (Kennedy, Yang, and Cohen 2010). Similarly, the Vegetation Change Tracker (VTR) is designed to track abrupt change utilizing thresholding (C. Huang et al. 2010). However, both approaches are univariate and unable to accommodate multiple metrics in a single analysis. The newest innovations in remote sensing utilize machine learning and multiple spectrally derived metrics (Moran, Kane, and Seielstad 2020; D’Este et al. 2021; Sun et al. 2022). Machine learning approaches can also integrate information from other sensors such as LiDAR.
LiDAR is well-suited for quantifying forest structure because it has capacity for fine spatial resolution and is the best sensor available for measuring in the vertical dimension (Hyde et al. 2006). The comparative advantage of LiDAR in machine learning is the massive training datasets it avails along with systematic and consistent measurements of forest structural attributes. However, LiDAR alone is limited as a monitoring tool because there is non-continuous spatial coverage and no built-in repeatability compared to passive, satellite based remote sensing systems such as Landsat and Sentinel. LiDAR-Landsat covariance is the process of identifying relationships between coincident Landsat imagery and LiDAR observations to overcome the limitations of each respective sensor. This method can be used to extrapolate LiDAR-derived metrics beyond their original geographic footprints (Hudak et al. 2002; Wilkes et al. 2015; Matasci, Hermosilla, Wulder, White, Coops, Hobart, and Zald 2018) or to update older LiDAR data without collecting another acquisition (Matasci, Hermosilla, Wulder, White, Coops, Hobart, Bolton, et al. 2018).
This study uses Lidar-Landsat covariance to produce a systematic, landscape-scale assessment of fire impact and forest recovery in the Northern Rocky Mountains. The purpose of the systematic, remote sensing approach is to contextualize the field-based findings of the regional studies cited previously. Gradient boosted regression models (GBMs) were developed to predict canopy cover from five spectral indices and three topographic variables. The models trained on forest cover measurements derived from ten LiDAR datasets encompassing 160,000 ha. Annual cover estimates were then produced for the period 1985–2021 at 30m spatial resolution across the Selway-Bitterroot and Bob Marshall wilderness complexes, to create 352 individual 35-year time series depicting the evolution of canopy cover on a census of burned sites.
The research builds on a legacy of remote sensing tools to evaluate forest impact and recovery from wildfire in the Northern Rocky Mountains. Time to recover from disturbance, expressed in years, has not been examined in these systems, although similar spectral-based approaches have been used to monitor boreal forests north of the study area (Schroeder et al. 2011; Bolton, Coops, and Wulder 2015; Pickell et al. 2016). The work’s novelty is in the integration of multiple sensors to inform a growing debate around the prognosis of forest recovery. It harnesses a training set of four million pixels to produce consistent cover estimates with known accuracy, and it can be updated easily to monitor into the future. Designated wilderness is well-suited for this investigation because its unique management context provides a natural laboratory for observing fire and vegetation dynamics relatively unfettered by human influences such as harvesting, planting, and reseeding (Kreider et al. 2022).
The objectives of this study are to:
-
Monitor canopy cover on 352 burned wilderness sites from 1985–2021 using machine learning to integrate forest observations from multiple sensors.
-
Quantify fire impact and whether or not burned areas are recovering canopy cover.
-
Estimate years to recovery based on post-fire trajectories of canopy cover.