Leveraging the Landsat archive to characterize plant species diversity and post-fire recovery in Great Basin shrublands

This masters thesis project had two chapters using remote sensing and spatial analysis of Great Basin rangelands, 1) remote sensing analysis and modeling of plant species diversity and 2) time-series analysis and modeling of post-fire vegetation recovery. Analysis was conducted primarily in R and Google Earth Engine.

  • Project lead Eric Jensen
  • Collaborators Jody Vogeler, Beth Newingham, Jason Sibold
  • Website Thesis document
  • Tools Google Earth Engine, R, Python, ArcGIS Pro
  • Completed October 2020


Project Summary

Great Basin shrublands are rapidly converting to plant systems dominated by noxious annual herbaceous plants, driven primarily by increased fire activity. Post-fire vegetation recovery trajectories vary spatially and temporally influenced by the effects of topography, climate, soils, and pre-fire vegetation. Our study leverages spatially continuous Landsat data alongside spatial models of environmental drivers, to account for variability across space and to evaluate important drivers of post-fire vegetation recovery. We first tested the spectral heterogeneity hypothesis, which suggests that variation in spectral values relates to plant species diversity, which, in turn, is theorized to be an important predictor of resilience to disturbance and resistance to invasive species. Weak relationships from the spectral heterogeneity tests led us to explicitly model species richness using both Landsat spectral data and environmental predictor variables. To evaluate drivers and patterns of post-fire vegetation recovery, we assessed pre- and post-fire trends of Landsat models of plant functional groups based on the number of times burned, post-fire seeding, and a suite of environmental predictor variables including pre-fire species richness. We also applied the suite of predictors to model vegetation recovery (15-year post-fire functional group dominance) and used the model to predict recovery for a contemporary fire, the Saddle Draw Fire from 2014. Our model of species richness had robust validation statistics with R2 value of 0.524, root mean square error of 6.69, and mean absolute error of 5.03 and evaluation of variable importance elucidated key drivers. While species richness may be important for aspects of ecological functioning not addressed in this study, it was not found among the most important drivers of post-fire recovery within Great Basin shrubland systems. What we did find to be a key driver of post-fire recovery was number of times burned which had a cumulative effect leading to increased annual herbaceous invasion and diminished perennial plant components. Meanwhile, on average post-fire seeding treatments had negligible influence upon post-fire perennial plant recovery. Post-fire recovery trajectories varied significantly across the fires evaluated in terms of both number of times burned and post-fire seeding. Models of post-fire recovery produced R2 values of 0.830 across fires evaluated and 0.338 when applied to new fires not included in model development. Spatially continuous analyses are important because they can account for spatial drivers of variability in post-fire recovery of Great Basin shrublands. While such analyses have previously been hampered by data limitations, our results suggest that advances in data availabilities and cloud computing resources may be increasing opportunities for adopting spatial approaches for providing ecological insight and to inform post-fire management decision making.

Recorded presentation

Key Figures

Species richness modeling and Landsat spectral heterogeneity analysis

We initially tested relationships between 94 unique measures spectral heterogeneity of Landsat TM, ETM+ and OLI, imagery and a plant species richness (using 10,471 field plots), finding weak but highly significant relationships at the broad extent of the Great Basin.

We then modeled plant species richness directly using random forest and XGBoost algorithms with 220 spectral, soils, fire, climate, and topography variables as candidate predictors. We used variable reduction to remove multicollinear and correlated variables and to select the most predictive layers, which produced and r-squared of 0.52 when applied to our validation dataset. The ten most predictive layers closely matched those found in similar field studies.

We produced predictive maps of plant species richness for Great Basin shrublands for each year from 1994-2017, finding distinct spatial pattern at broad scales that coincide with understood patterns of plant invasion and altered fire regimes.

Post-fire vegetation recovery analysis and modeling

We evaluated the role of the repeated fire and post-fire management on vegetation recovery trajectories using the Normalized Differenced Perennial Dominance Index (NDPDI) as a response variable. We found that repeated fire had a cumulative affect of diminished post-fire recovery overall, but with variability across the region. In terms of seeding, no appreciable improvement in post-fire recovery was found as a result of seeding when compared to unseeded areas.

We then used vegetation,soils, fire, climate, management, and topography variables to explicitly model post-fire vegetation recovery 15-years following fire using random forest modeling, selecting 17 variable for the final model. We validated the model using conventional methods (r-squared of 0.83) and using leave-one-out validation (mean r-squared of 0.34). Overall, pre-fire vegetation, climate aridity, and southern aspects and high heat load indices played driving roles in post-fire vegetation outcomes.

The final analysis focused on predicting post-fire recovery patterns for a contemporary fire, the Saddle Draw Fire which burned 284,065 acres in southeastern Oregon in 2014. We produced a predictive map of post-fire recovery depicting NDPDI 15-years following the fire.