| 19 Nov 2025
Modelling N₂O Emissions in Fertilized Grasslands with Machine Learning: A Multi-Site LSTM Approach
Abstract. Accurate quantification of GHG emissions from agricultural soils based on key environmental and management factors, is crucial for developing effective mitigation strategies. This study applies a machine learning approach using data from multiple montane grasslands in central Europe to model the spatial and temporal dynamics of N2O emissions, using meteorological, soil, and management data. The primary aim is to advance predictive modelling of N2O emissions from grassland soils for spatial upscaling and scenario analysis. Specifically, we assess whether a generic model can accurately estimate N₂O emissions from an independent site excluded from training.
We collected data from five fertilized grasslands across southern Germany, northern Switzerland and northern Austria (122 soil-site-treatment-year combinations), and trained a long short-term memory (LSTM) algorithm to model the influence of drivers within 5 subsequent days on N2O emissions. The dataset includes daily N2O emission measurements along with key emission drivers such as soil moisture and temperature in 10 cm soil depth, daily precipitation, occurrence of fertilization events (zero or one flag), as well as soil characteristics such as pH and bulk density.
The trained LSTM model showed strong predictive performance (RMSE of 18 μgm−2h−1, and Relative RMSE of 270 %) when evaluated on a test set that included both data from an independent soil and withheld years from training procedures. The model accurately captured N2O dynamics, including the magnitude and timing of emission peaks driven by slurry application and environmental factors. Compared to the performance of established process-based biogeochemical models the LSTM model yielded similar RMSE and bias values for most site-years. These results demonstrate that LSTM-based models can reliably predict N₂O emissions at independent sites with similar environmental and soil characteristics and represent a promising alternative to process-based models for predicting soil N2O emissions.
