Making full use of hyperspectral data for gross primary productivity estimation with multivariate regression: Mechanistic insights from observations and process-based simulations

Abstract

Statistical gross primary productivity (GPP) estimation from remote sensing observations has mostly been attempted on the basis of multispectral observations. To make full use of the information contained in vegetation spectra, however, hyperspectral observations should be used in combination with appropriate multivariate methods. Nevertheless, only very few previous studies attempted to estimate GPP directly from hyperspectral observations and did so on the basis of reflectance, with only a limited number of temporally discontinuous observations. In this study, we used long-term, continuous, half-hourly hyperspectral observations covering the visible and near-infrared spectral range to estimate GPP directly from upwelling irradiance using partial least square (PLS) regression in a rice paddy. To gain a better understanding of processes underlying the PLS estimation, we used extensive complementary field observations to run process-based simulations using the SCOPE model. We then applied PLS regression to the simulated hyperspectral data in the same way as for the observations and disentangled contributions related to relevant physiological processes, namely sun-induced chlorophyll fluorescence (SIF) and xanthophyll cycle-related spectral changes (XC). We found that upwelling hyperspectral irradiance in the visible and near-infrared spectral range predicted GPP better than reflectance. Furthermore, PLS-based GPP estimates outperformed both far-red SIF and widely used vegetation index-based methods. However, the most relevant information for the observation-based PLS-models was not clearly related to XC or SIF as the near-infrared spectral range showed comparable performance. Also, the simple average of upwelling irradiance over the 850–900 nm range outperformed the other non-multivariate approaches, including far-red SIF. These results held for the evaluation in terms of the seasonal variation of GPP, while there was apparently a small contribution of SIF and XC for the diurnal variation. The simulation-based analysis showed that SIF and XC contributed useful information to both GPP and photosynthetic light use efficiency (LUE) estimates at both seasonal and diurnal time scales. The strongest unique contribution from either SIF or XC, however, was to the diurnal variation of GPP and XC showed considerably better performance than SIF. We did not find improvements when combining the spectral regions of XC (500–570 nm) and SIF (650–800 nm) to estimate GPP. While SIF showed improvements when combined with the remaining spectral information excluding XC, this was not the case for XC. Our approach combines the strengths of process-based modeling with multivariate statistical analysis to improve our understanding of the usable information content in vegetation spectra and is highly relevant for further developing suitable methods for GPP estimation at large scales.