Retrieval of Arctic Vegetation Biophysical and Biochemical Properties from CHRIS/PROBA Multi-Angle Imagery Using Empirical and Physical Modelling

Abstract

Mapping and monitoring of Arctic vegetation biochemical and biophysical properties is gaining importance as global climate change is disproportionately affecting this region. Previous studies using remote sensing to model Arctic vegetation biochemical and biophysical properties have generally involved empirical modelling with nadir looking broadband sensors and have typically been conducted at the field scale in one study area. Satellite hyperspectral remote sensing has not been previously investigated for retrieving leaf and canopy biochemical and biophysical properties of Arctic vegetation across multiple sites using either empirical or physically-based modelling approaches. Furthermore, multi-angle hyperspectral sensors (CHRIS/PROBA), which can provide insight into vegetation reflectance anisotropy and potentially improve vegetation parameter estimation, have also not been investigated for this purpose. In this study, three modelling approaches previously investigated with field spectroscopy data (Kennedy et al., 2020) were used with CHRIS Mode-1 imagery to predict leaf chlorophyll content, plant area index and canopy chlorophyll content across a bioclimatic gradient in the Western Canadian Arctic. Modelling approaches included: parametric linear regression based on vegetation indices (VI), non-parametric machine learning Gaussian processes regression (GPR) and inversion of the PROSAIL radiative transfer model using a look-up table approach (LUT). CHRIS imagery was acquired with −55°, −36°, 0°, +36°, +55° view zenith angles (VZA) between 2011 and 2014 over three field sites extending from the Richardson Mountains in central Yukon, Canada to the north end of Banks Island, Northwest Territories, Canada. Field measurements were acquired within several weeks of satellite acquisitions. GPR had the best model fit (mean cross-validated (cv) coefficient of determination, r2cv = 0.61 across all vegetation variables, sites and VZAs vs. 0.59 for the simple ratio, SR) and predictive performance (normalized root mean square error, NRMSEcv = 0.13 vs. 0.14 for SR). The revised optimized soil adjusted VI (ROSAVI) performance was slightly poorer (r2cv = 0.51; NRMSEcv = 0.15). The physically-based PROSAIL model performed poorer than all empirical models (r2 = 0.50; NRMSE = 0.18). This ranking of model performance is similar to that found in the previous field spectroscopy study, where empirical model fits and predictive performance were only slightly worse. With respect to view angle performance, NRMSE varied only slightly, indicating no distinct advantage for any one VZA. Overall, strong potential has been demonstrated for empirical modelling of Arctic vegetation chlorophyll and plant area index using hyperspectral data combined with band selection/optimization procedures in the Arctic. Recently launched and future hyperspectral satellites, including next generation airborne sensors, will likely provide improvements to the model performance reported here.