Estimating global transpiration from TROPOMI SIF with angular normalization and separation for sunlit and shaded leaves

Abstract

Gross primary productivity (GPP) is more accurately estimated by total canopy solar-induced chlorophyll fluorescence (${\mathit{SIF}}_{\mathit{total}}$${\mathit{SIF}}_{\mathit{total}}$) compared to raw sensor observed SIF signals (${\mathit{SIF}}_{\mathit{obs}}$${\mathit{SIF}}_{\mathit{obs}}$). The use of two-leaf strategy, which distinguishes between SIF from sunlit (${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$) and shaded (${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$) leaves, further improves GPP estimates. However, the two-leaf strategy, along with SIF corrections for bidirectional effects, has not been applied to transpiration (T) estimation. In this study, we used the angular normalization method to correct the bidirectional effects and separate ${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$ and ${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$. Then we developed ${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$ and ${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$ driven semi-mechanistic and hybrid models, comparing their T estimates with those from a ${\mathit{SIF}}_{\mathit{obs}}$${\mathit{SIF}}_{\mathit{obs}}$ driven semi-mechanistic model at both site and global scales. All three types of SIF-driven T models integrate canopy conductance ($g_c$$g_c$) with the Penman-Monteith model, differing in how $g_c$$g_c$ is derived: from a ${\mathit{SIF}}_{\mathit{obs}}$${\mathit{SIF}}_{\mathit{obs}}$ driven semi-mechanistic equation, a ${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$ and ${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$ driven semi-mechanistic equation, and a ${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$ and ${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$ driven machine learning model. When evaluated against partitioned T using the underlying water use efficiency method at 72 eddy covariance sites and two global T remote sensing products, a consistent pattern emerged: ${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$ and ${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$ driven hybrid model > ${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$ and ${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$ driven semi-mechanistic model > ${\mathit{SIF}}_{\mathit{obs}}$${\mathit{SIF}}_{\mathit{obs}}$ driven semi-mechanistic model. The ${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$ and ${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$ driven hybrid model demonstrated a notable proficiency under high vapor pressure deficit and low soil water content conditions. The ${\mathit{SIF}}_{\mathit{obs}}$${\mathit{SIF}}_{\mathit{obs}}$ driven semi-mechanistic model tends overestimate T at low T values, and this issue is significantly alleviated by the ${\mathit{SIF}}_{\mathit{sunlit}}$${\mathit{SIF}}_{\mathit{sunlit}}$ and ${\mathit{SIF}}_{\mathit{shaded}}$${\mathit{SIF}}_{\mathit{shaded}}$ driven semi-mechanistic and hybrid models. Our findings demonstrate that correcting the bidirectional effects and using the two-leaf strategy on GPP estimation can improve T estimation and provide a new global T product incorporating vegetation physiological signal.