GROUNDED EO: Data-driven Sentinel-2 LAI and FAPAR retrieval using Gaussian processes trained with extensive fiducial reference measurements

Abstract

Due to their importance in monitoring and modelling Earth's climate, the Global Climate Observing System (GCOS) designates leaf area index (LAI) and the fraction of absorbed photosynthetically active radiation (FAPAR) as essential climate variables (ECVs). The Simplified Level 2 Biophysical Processor (SL2P) has proven particularly popular for decametric (i.e. 10 m to 100 m) retrieval of these ECVs. Comprehensive validation has shown that due to simplifying assumptions in the underlying radiative transfer models (RTMs), biases persist in SL2P retrievals. To avoid RTM assumptions altogether, an empirical data-driven approach might be considered. Yet, such a strategy has historically been prevented by the limited quantity and quality of available in situ reference measurements, as well as the large number of training samples traditionally required by machine learning regression algorithms. New opportunities are now offered by recently established continental-scale environmental monitoring networks, advances in automated data processing and uncertainty evaluation, and machine learning regression algorithms that require many fewer training samples. The Ground Reference Observations Underlying Novel Decametric Vegetation Data Products from Earth Observation (GROUNDED EO) project was initiated to take advantage of these opportunities. We describe the empirical data-driven LAI and FAPAR retrieval approach adopted within the project, involving i) generation of a database containing over 16,000 fiducial reference measurements covering 81 National Ecological Observatory Network (NEON), Integrated Carbon Observation System (ICOS), and Terrestrial Ecosystem Research Network (TERN) sites between 2013 and 2022, ii) development of an empirical data-driven algorithm for Sentinel-2 LAI and FAPAR retrieval based on Gaussian processes, and iii) evaluation of GROUNDED EO retrievals through intercomparison with the current state-of-the-art in decametric retrieval (i.e. SL2P, and a modified version of SL2P developed by the Canada Centre for Remote Sensing – SL2P-CCRS), as well as validation against unseen fiducial reference measurements. In the majority of cases (and despite not making use of ancillary data such as land cover), the empirical data-driven GROUNDED EO retrievals were subject to reduced bias than those from SL2P and SL2P-CCRS, as well as increased fulfilment of user requirements (i.e. 74% of LAI and 69% of FAPAR retrievals overall). Consequently, the approach has potential to reduce uncertainty in key inputs for climate monitoring and modelling, agricultural and forest management, and biodiversity assessment.