Shrub willow canopy chlorophyll content estimation from unmanned aerial systems (UAS) data: Estimation and uncertainty analysis across time, space, and scales

Abstract

Remote sensing is an important technology to map land-surface parameters, with many studies demonstrating land-surface parameter estimation, but fewer testing model transferability and quantifying model parameter uncertainty. In this study, we evaluated the uncertainty across time, space, and spatial scales of shrub willow health characterization based on canopy chlorophyll content (CCC) derived from unmanned aerial systems (UAS) data. Shrub willow is a short-rotation woody crop used to produce biomass for bioproducts and CCC is a popular biophysical parameter used to indicate shrub willow health status. We analyzed time-series UAS data at three spatial scales (5 m, 10 m, and 20 m) and stratified field-observed CCC levels. Since scale effects are often related to spatial heterogeneity, we implemented a nested analysis of variance (ANOVA) to evaluate the spatial heterogeneity within different pixel sizes and found that 5 m pixels were the most homogeneous, followed by 10 m and 20 m. Results from regression modeling of shrub willow CCC as a function of red-edge normalized difference vegetation index (NDVIre) at 5 m, 10 m, and 20 m scales showed that the models built at 5 m, 10 m, and 20 m could be applied across time, space, and scales. We also quantified the uncertainty for model parameters using two different inferential frameworks, confidence interval (CI) widths from the frequentist framework and credible interval (CrI) widths from the Bayesian framework. Unlike predictive root mean square error (RMSE), which showed model output uncertainty decreased as pixel size increased, CI and CrI widths indicated that the related model parameter uncertainty increased as pixel size increased. We calculated CI and CrI widths based on time-series model building and different stratified model building to analyze model parameter uncertainty and quantified predictive RMSE for all sampling date combinations. The results showed that, in terms of model complexity and the range of ground observations, both frequentist and Bayesian regression have advantages and disadvantages and demonstrated that the uncertainty quantified by CrI widths is able to guide future experimental design to save resources.