Long-term climatic and management drivers of cropland gross primary production (GPP) from eddy covariance explained by interpretable machine learning

Abstract. Understanding the long-term response of cropland gross primary production (GPP) to climate variability remains limited, despite the key role of agricultural carbon fluxes in climate change mitigation. Here, we analyze two decades of eddy covariance measurements from a Swiss cropland site to assess how climate and management influence GPP dynamics of three winter crops: wheat, barley, and rapeseed. Crop-specific growing seasons were characterized by fitting logistic functions to cumulative GPP as a function of growing degree days, allowing the extraction of phenological metrics and growth rates. In parallel, an eXtreme Gradient Boosting (XGBoost) model combined with SHapley Additive Explanations (SHAP) was used to identify the dominant climatic and management drivers of daily GPP and their temporal patterns.

Air temperature and vapor pressure deficit exhibited significant increasing trends over the study period, consistent with ongoing climate change. However, no corresponding trends were detected during either actual or fixed winter growing season (October–July). Instead, significant warming and atmospheric drying were determined to occur during the post-harvest off-season (August–September). This shows that a crop rotation dominated by winter crops acts as a climate-smart strategy for the Oensingen site, enabling crops to escape intensifying peak summer heat and atmospheric dryness. Across crops, cumulative GPP trajectories and growth rates were largely similar. Winter wheat was the only crop showing a significant increase in maximum cumulative GPP over time, yet this increase was not reflected in grain yield, confirming that GPP is a poor predictor of harvested production. Machine-learning results identified incoming radiation, air temperature, and nitrogen fertilization as the dominant drivers of GPP, highlighting the importance of explicitly accounting for management variables in cropland carbon flux studies.

Overall, the absence of strong crop-specific divergence and limiting responses suggests that the studied crops have not yet experienced climatically constraining conditions for carbon uptake. These findings emphasize the need for long-term, multi-site cropland observations that integrate management data to better assess crop-climate interactions under future climate change.