Rivers aren’t static and square, so why model them like they are? Developing a Bayesian hierarchical model of LiDAR derived riparian canopy height on the Yuba River

Background/Question/Methods

Previously, elevation, depth to groundwater, inundation frequency, and substrate were found to be strong deterministic predictors of canopy composition and height in riparian corridors. Today, high resolution (0.914 x 0.914 m) LiDAR provides detailed, spatially explicit canopy datasets to test both deterministic and stochastic contributions to riparian spatial structure. We collected LiDAR and other data in a ~28.3 km stretch of the semiarid lower Yuba River, CA and then developed a hierarchical Bayesian model of canopy height, conditioned by covariates based on positioning. Relative to the river, these covariates followed four different spatial gradients: 1) streamwise, 2) lateral, 3) forced by local river features or topography, and 4) independent of river landforms.  Introducing zero-centered noise to the model allows for unrepresented covariates to be included. Modeling the spatially-lagged autoregression mitigated the effects of autocorrelation within the data. Preliminary work required the assessment of a prioridistributions and metrics of spatial autocorrelation.

Results/Conclusions

Before we could build a conditional model, we assessed the distributions of canopy data stratified by landform type using beanplots. The canopy height LiDAR data were log-normally distributed. The canopy, dominated by willows, was lower to the ground on in-channel features and bars subjected to frequent inundation. This was consistent with a general trend of greater canopy heights further from the river quarter, excepting areas of human disturbance (e.g., levees). Canopy height along the river bank was directly related to the distance downstream. This relationship was far weaker in regions of less frequent inundation: farther from the channel. This suggests that the effects of relative streamwise position (1 above) on canopy height be modeled as conditional upon the position perpendicular to the flow (2 above). Moran’s I is a metric of spatial autocorrelation within data, which varies on a scale of -1 (negatively autocorrelated) to 1 (positively autocorrelated). The LiDAR derived canopy data were found to be substantially autocorrelated (Moran’s I = 0.95). Given the incredibly high resolution of the data, and the large-spatial scales at play with many of the covariates, we expected the canopy-height data to be very strongly and positively autocorrelated.