An end-to-end satellite-based GPP estimation model devoid of meteorological and land cover data

Abstract

Vegetation gross primary productivity (GPP) plays a key role in the terrestrial carbon cycle, and remote sensing models are one of the main tools for GPP estimation at regional and global scales. Almost all existing remote sensing models (e.g., regression models, parametric models, process models and machine learning models) rely on plant functional type (PFT)-based parameter settings, multiple data sources (e.g., meteorological data) and key indices (e.g., leaf area index, LAI), limiting their estimation accuracy and spatial generalization capability. Therefore, we developed an End-To-End Satellite-based model (ETES) to improve GPP estimation. ETES only utilizes input variables from original satellite observations and Global Land Surface Satellite (GLASS) downward shortwave radiation data. It replaces the traditional vegetation types data with a set of numeric variables (named as Seasonal Characteristics of Vegetation Types and Growth, SCVTG) derived from the curve of vegetation index time series within each growing cycle. The multi-layer perceptron method was applied to model the end-to-end relationship between GPP and input variables. Taking the flux data from FLUXNET 2015 as the benchmark, the GPP estimation accuracy of ETES was higher than that of similar GPP products (i.e., MOD17, GOSIF, GPP-NIRv and FLUXCOM RS), with an average 27.89% reduction in RMSE (ΔRMSE, -0.96 ∼ -0.6 g C m−2 day−1) and a 28.86% increase in R2 (ΔR2, 0.09 ∼ 0.22) at the monthly scale. In short, ETES can effectively improve GPP estimation in data availability, spatial generalization capability and estimation accuracy. Meanwhile, SCVTG, as a new proxy of real vegetation types and phenology, would benefit the design of terrestrial carbon flux estimation models.