eLTER DAR

Persistent record identifiers

DOI - 10.23728/b2share.41d638b6c1534057b71abcca170ab7d0

Data keywords

Projects

B2SHARE external record - eLTER Community

Site References

Site reference:

Mejusseaume - France

Creators

Anna Spinosa

Given Name: Anna

Family Name: Spinosa

Mario Alberto Fruentes-Monjaraz

Given Name: Mario Alberto

Family Name: Fruentes-Monjaraz

Valeria Mobilia

Given Name: Valeria

Family Name: Mobilia

Abstract

Gross Primary Production (GPP) represents the total amount of carbon fixed by plants through photosynthesis in an ecosystem over a specific period. GPP data products are derived using a data-driven approach that integrates Earth observation data with in-situ carbon flux measurements. Specifically, GPP estimations combine Sentinel-2 multispectral imagery with carbon flux data from eddy covariance towers, employing the XGBoost machine learning algorithm for prediction. The resulting GPP maps are generated at a 10-meter spatial resolution and a temporal frequency up to 5 days, covering the period from March 2017 to December 2023. The temporal resolution is contingent upon 50% free-cloud conditions in the area of interest, with lower frequencies occurring during periods of high cloud coverage. The spatial extent of the GPP maps corresponds to the boundaries of long-term observation sites as recorded in the DEIMS-SDR registry (e.g., https://deims.org/a5496211-d63f-494b-8da7-64e9abf8898b). For sites where the boundary area is smaller than 1 km², or if only point coordinates are available in DEIMS-SDR, the maps are constrained to a 1 km x 1 km bounding box.

Description - Methods

The methodology integrates multiple data sources via machine learning techniques to estimate Gross Primary Production (GPP) across different ecosystems. The process begins with data pre-processing, including the selection of sites based on criteria such available vegetation information, at least three full years of eddy covariance flux data. GPP and environmental data (e.g. air temperature, vapor pressure deficit, etc.) are extracted from the ICOS database across different ecosystem types. Then, different remote sensing (RS) indices (e.g NDVI, EVI, etc.) are estimated in GEE using Sentinel-2 MSI data as the mean value of the pixels found inside the climatological footprint 70 (an area were 70% of the GPP measurements are coming from). Both data coming from ICOS dataset and RS indices are used as predictors for the model. The data is split, with 70% used for training and 30% for testing. In the model setup, an XGBoost model is trained using the selected environmental and RS based indices. The model parameters are fine-tuned to improve accuracy. The remaining 30% of testing data is used to evaluate the model’s performance by comparing its predictions against in-situ GPP data. Error metrics like Mean Absolute Error (MAE), Root Mean Absolute Error (RMAE), and R² are provided. The maps computation phase applies the trained model to ecosystem boundaries from the DEIMS website to generate 5-day GPP maps.

Description - Other

Acknowledgement This work on the AGAME Gross Primary Production data product is funded by the European Space Agency (ESA, contract no. 4000143740/24/I-AG) in the frame of the GEOSS Platform Plus project (Horizon Europe, GA No. GA.Nr. 101039118). The work done is based on the requirements from eLTER contributing in addition to the eLTER Site Information Cluster. In-situ data for model calibration and validation has been derived from the ICOS Carbon Portal.

Licences

cc-by

External Source Information

B2Share

Additional Metadata

publication_state: published

community: d952913c-451e-4b5c-817e-d578dc8a4469

External files

File format: pdf

1048447 kB

Checksum

File format: zip

3797280 kB

Checksum

File format: csv

236 kB

Checksum

File format: zip

103575540 kB

Checksum

Online services and data

Mejusseaume - Gross Primary Production (2020-2023)

Site reference:

B2Share