2.5. Integration of EBD-GPR models in GEE

The GEE data catalog provides a multi-petabyte collection of widely used satellite imagery, including the complete archives of S2 L1C TOA and L2A orthorectified atmospherically corrected surface reflectance data. On the processing side, the Python Application Programming Interface (API) package ee provides functions that allow to extract any available information layer over a specific area of interest (AOI) and process the resulting datasets very efficiently, thus enabling studies at any place on Earth and any time since the launch of Sentinel-2A in 2015. In GEE, for L1C and L2A the datasets are available from 23 June 2015 and 28 March 2017, respectively.

Additionally, the ee library provides optimized functions to perform computationally expensive cartographic (mosaicking, compositing, clipping, etc.) and matrix algebraic operations. To circumvent time/space bottlenecks and fully exploit the high-performance parallel computing environment, all these operations must be carried out on server-side of Google processing facilities (Pipia et al., 2021). All in all, a whole study can be bounded to the AOI starting off with the raw images from the available datasets in a few steps. As a demonstration case, two S2 tiles (T32UPU and T32UQU) acquired over the MNI study site on 6 July 2017 were selected. First, the corresponding 10-to-20 m bands were mosaicked and then clipped over the AOI. The maps of the different functional vegetation traits were obtained by importing the corresponding EBD-GPR models generated in ARTMO in the GEE environment and performing the mean value prediction on-the-fly as explained in Pipia et al. (2021). Essentially, the standard formulation of anisotropic-kernel GPR in Camps-Valls et al. (2018) was reorganized in a way to isolate all those terms depending on the model’s hyperparameters and training data, which can be calculated before being imported into GEE. The remaining ones, which account for the mathematical bounds between the new input (i.e., the multispectral S2 imagery to be processed) and specific features of the GPR model, are decomposed into matrix linear algebra operations, which are suited for parallel implementation into GEE. As already mentioned, this computational optimization can be achieved if only functions provided by the ee library are used for coding. As a result, the GPR mean value retrievals from a specific S2 tile can be visualized in a few seconds at any zoom level, and the maximum resolution map can be downloaded locally in a few minutes. In Section 3.3, the process is applied to all the S2 TOA EBD-GPR models corresponding to the multiple crop traits over the selected AOI, and in Section 3.4 the resulting maps are compared against the corresponding SNAP estimates. Finally, Section 3.5 shows an example of how to exploit GEE to the fullest, where all the vegetation traits described in Table 2 are mapped at the country scale of Germany. Since mapping the country scale may not fully reveal the details of the obtained trait maps, a few subset maps were additionally generated zooming into specific agricultural regions of Germany at 20 m GSD (see Section 3.5). The GEE codes to run the EBD-GPR models and display the vegetation maps of this study is hosted on the repository https://github.com/esjoal/GEE_GPR_mapping_vegetation.