Physics-Aware Machine Learning for Carbon Fluxes at High Spatio-Temporal Resolution and Scales

Carbon captured via photosynthesis by vegetation is known as gross primary production (GPP). It is an important variable related to climate regulation and determines ecosystem carbon sources and sinks. GPP is routinely estimated globally by operational algorithms that combine remote sensing data at coarse spatial scales (e.g., MODIS, 500 m) and meteorological information. The need for global high-resolution operational products arises from the requirement of capturing GPP variability, which co-occurs at finer resolutions and over large areas. These specifications demand gap-free remote sensing data to obtain continuous maps, high spatial and temporal resolution, and realistic uncertainty quantification. Machine learning (ML) methods are widely used but sometimes do not fit real-world physics restrictions. Therefore, we propose a physics-aware machine learning methodology that combines 1) high spatial resolution spectra at 30m and gap-free observations derived from blending Landsat and MODIS with the HISTARFM algorithm, 2) meteorological information, and 3) in situ eddy covariance GPP estimates as reference data. The ML model further incorporates an extra regularizer that constrains the GPP estimates for improved consistency with ancillary data and covariates closely related to photosynthesis (e.g., SIF). Moreover, we rely on the HISTARFM methodology to provide well-calibrated data uncertainty estimates, which allows us to yield both epistemic and aleatoric uncertainty for the GPP estimates. The processing pipeline is fully implemented in Google Earth Engine (GEE), allowing us to estimate carbon fluxes over Europe at 30m. The methodology enables more precise and real-world carbon studies and opens the door to deriving other key fluxes at an unprecedented spatiotemporal resolution.