Quantifying agricultural traits and land surface phenology metrics in Google Earth Engine.

Monitoring of crop growth, variability and dynamics over agricultural areas is needed to optimize management practices and thus to ensure global food security. Nonetheless, estimation of cropland phenology from optical satellite data remains a challenging task due to the influence of clouds and atmospheric artifacts. Therefore, measures need to be taken to overcome these challenges and gain better knowledge of crop dynamics. 

Since 2017, the European Space Agency (ESA) Copernicus Sentinel-2A & B (S2) have been providing high resolution optical imagery all over the globe with an observation frequency of 5 days. With 13 spectral channels and 10-60m spatial resolution, time series of these data offer untapped potential for monitoring cultivated areas. In this respect, the processing of S2 imagery in cloud-based platforms, such as Google Earth Engine (GEE), allows large-scale precise mapping of agricultural fields. The arrival of GEE enabled us to propose an end-to-end processing chain for vegetation phenology characterization using S2 imagery at large scale.

To achieve this, the following pipeline was implemented: (1) building hybrid Gaussian process regression (GPR) models optimized with active learning (AL) for retrieval of crop traits, such as leaf area index (LAI), fractional vegetation cover (FVC), canopy chlorophyll content (laiCab), canopy dry matter content (laiCm) and canopy water content (laiCw), (2) implementing these models into GEE, (3) generating spatially continuous maps and gap-filled time series of these crop traits, and finally (4) calculating land surface phenology (LSP) metrics, such as start of season (SOS) or end of season (EOS), by using the conventional double logistic approach.

In respect to step (1): variable-specific training datasets were generated in the ARTMO software environment using PROSAIL model simulations, with training samples reduced in number but optimized in quality, i.e. representativeness, using the Euclidean-distance based (EBD) AL technique. In this way, light retrieval models were generated via GPR, a ML algorithm which builds up a retrieval model by learning the non-linear relationships between the spectral signals and crop traits of interest. Overall, good to high performance was achieved in particular for the estimation of canopy-level traits, such as LAI and laiCab, with normalized root mean square errors (NRMSE) of 9% and 10%, respectively. Subsequently, (2) the retrieval models were integrated into the GEE environment to perform mean value prediction on-the-fly. In this way, time series of crop traits based on S2 images were produced quasi-instantly over the area of interest. As demonstration of the workflow capability to easily reconstruct time series of S2 entire tiles, phenology maps from multiple crop traits were generated over an agricultural area in Castile and Leon, Spain. For this region also crop calendar data were available to assess the validity of the LSP metrics derived from crop traits. In addition, LSP metrics derived from the Normalized Difference Vegetation Index (NDVI) were used as reference, demonstrating the good quality of the quantitative traits products to describe phenology. Thanks to the GEE framework, the proposed workflow can be carried out globally in any time window, thus representing a shift in satellite data processing towards cloud computing.