LiDAR-Landsat Covariance for Predicting Canopy Fuels

Managing wildfires in the western United States is becoming increasingly complex. Visualizing and quantifying canopy structures allows fire managers to both plan for fire and track recovery. Light detecting and ranging, or LiDAR can measure forests in three dimensions, but has limited spatial and temporal coverage. LiDAR-Landsat covariance uses machine learning to fill in the spatial and temporal gaps of LiDAR coverage with supplemental Landsat imagery. However, in order to capture real forest dynamics, a model needs to be stable enough to detect long term trends, sensitive to episodic disturbance, and general enough to work on multiple landcovers. The purpose of this research is to refine the methodology behind LiDAR-Landsat covariance and assess if these predictions can yield sable and ecologically sensible time series to track forest fire recovery over time. Gradient boosted machine models (GBMs) were built to predict canopy cover, height, and base height. Then, they were tested on a series of validation sites in order to quantify the spatial and temporal sources of error associated with these models. Finally, the models were used to predict the trajectories of canopy cover, height, and base height on 164 fire scars in Montana, Idaho and Wyoming over the course of 36 years. The models were sensitive to moderate and high severity disturbance, both on an incident wide and pixel by pixel basis. Overall model R2 values were 0.89 for canopy cover, 0.84 for height, and 0.88 for base height. Year to year variability in canopy cover on validation sites was 2.3%. Height had more variability due to a sensor artifact from the transition from Landsat 5 to Landsat 8. On the Lost Fire the model found high severity fire corresponded with greater canopy fuel losses on a pixelwise basis. The models also detected canopy recovery, and found four distinct trajectories in which burned sites recover from disturbance. Seventy-seven percent of sites fully recovered canopy cover to pre-fire conditions within the 36-year time series. Further refinement of GBM based LiDAR-Landsat covariance can increase the sensitivity to smaller disturbances and reduce the impact of model error on performance