Automated Gross Primary Production Application for Monitoring Ecosystem Health

Providing quality controlled and up-to-date data is an important prerequisite for informed environmental policies and the implementation of management decisions especially on local and regional scale. Therefore, in the scope of the AGAME project we selected gross primary production (GPP) as a relevant ecosystem health indicator and future eLTER Standard Observation variable to produce data products that enable the monitoring and assessment of impacts of global change on ecosystems. Consistent GPP data products together with detailed metadata were co-designed with users and based on their requirements generated and published online. These co-designed data products encompass GPP maps produced for 16 different sites in Europe covering diverse ecosystems. The GPP data products were calculated with a data-driven approach combining Earth Observations with in-situ carbon data. GPP data products are provided at a 10 m spatial resolution generally for every 5 days for the period from June 2017 until December 2023. The period differs per site depending on the availability of in-situ data in the Integrated Carbon Observation System (ICOS). The highest temporal resolution of 5 days is achieved in periods with 50% free-cloud conditions in the area of interest. When cloud cover exceeds this threshold the data is excluded resulting in reduced frequency. An XGBoost model was trained and evaluated using Mean Absolute Error (MAE), Root Mean Absolute Error (RMAE) and coefficient of determination (R²) to predict the time series of GPP. Annual error metrics are provided to ensure model accuracy. Specifically, Sentinel-2 multispectral data is combined with environmental data derived from ICOS eddy covariance tower systems using the XGBoost (Chen and Guestrin 2016) machine learning algorithm (Spinosa et al. 2023) to generate GeoTIFF and NetCDF files. The map extent of each map corresponds to the boundaries of the selected long-term observation facilities as provided on the site registry DEIMS-SDR. If the site boundaries cover an area smaller than 1 km² or the site boundaries are not available in DEIMS-SDR with only point coordinates provided, the map extent is defined as a 1km x 1km bounding box to ensure consistency across all sites and balance computational efficiency with data availability. This extension has also been discussed with users engaged in the co-definition phase of the project. The final data products were published in the eLTER Cyberinfrastructure, B2Share and the GEOSS portal. The publication on B2Share ensured that each data product was stored persistently with dedicated DOIs. The generated GeoTIFFs were published in the eLTER Cyberinfrastructure using a dedicated cloud file storage system and GeoServer creating geospatial data services (WMS/WFS) that enable on-the-fly work with the time-series data in the browser or desktop GIS applications. Metadata records were created and exposed using the Catalogue Service of the Web (CSW) and linked to the GEOSS portal further increasing the findability and accessibility of the data products.