• Machine learning upscaling of static chamber CO 2 fluxes mirrored eddy-covariance estimates. • CO 2 flux modelling allowed to estimate management effects on grassland annual C balance. • Grazing exclusion and legumes overseeding reduced net C losses in a semi-arid grassland. • Drought increased net CO 2 emissions in a semi-arid grassland. Assessing carbon (C) fluxes in semi-arid grasslands is essential for designing adaptive management strategies in the face of climate change. However, continuous monitoring of CO₂ exchange using eddy-covariance (EC) systems is costly and often unsuitable for comparing treatments at small spatial scales. Upscaling portable chamber measurements using machine-learning models may offer a promising, yet still under-explored, alternative. This work evaluates the potential of this approach to assess the effects of grazing exclusion, rotational grazing and legume overseeding on grassland annual C balance, compared to conventional continuous grazing. Over two years, 1009 measurements of net ecosystem exchange (NEE), gross primary productivity (GPP), and ecosystem respiration (Reco) were collected using portable chambers at monthly to bi-monthly intervals across 16 paddocks with contrasting management in a Mediterranean silvopastoral system. Cubist regression models were trained on those observations using reanalysis climate data (ERA5, CAMS, SMAP), high-resolution normalized difference vegetation index (NDVI) from Sentinel-2 imagery, soil properties measured in situ and management categories as predictors. Modelled fluxes over a seven-year period were compared to EC continuous measurements at one of the paddocks. Upscaling chamber-based measurements produced reliable predictions of C fluxes at hourly, daily and annual scales. These fluxes were primarily driven by climatic conditions, with soil water content playing a central role. NDVI and soil fertility were also key predictors of C fluxes and mediated part of the management effects. Studied grasslands acted as net C sources, particularly during drier years. Grazing exclusion and recent (10 years) legume overseeding reduced C losses by 21%, 9%, and 22%, respectively, while rotational grazing had no effect on annual NEE budget, compared to continuous grazing. Management effects were independent of the inter-annual climate variability. This highlights legume overseeding as a resilient management alternative to improve C sequestration while maintaining productive semi-grasslands.
Carrascosa et al. (2026) studied this question.