Modeling carbonyl sulfide and carbon dioxide fluxes in a northern boreal coniferous forest using memory-based deep learning

Boreal forests are vital to the global carbon cycle but are highly sensitive to climate change. Accurate monitoring of their carbon dynamics is of great significance for effective climate change mitigation. However, modeling boreal forest carbon fluxes remains fraught with uncertainties. In this study, we investigated the potential of Long Short-Term Memory networks in predicting the fluxes of carbonyl sulfide (COS) and carbon dioxide (CO₂), and systematically analyzed the controlling factors of these fluxes across multiple time scales using eddy-covariance flux data from a coniferous forest in northern Finland (2013–2017). The results reveal that the model’s predictive performance for COS, Gross Primary Productivity (GPP), and Net Ecosystem Exchange (NEE) varies considerably across different time scales. At the half-hourly scale, the determination coefficients of the model for predicting COS, GPP, and NEE are 0.51, 0.62, and 0.75 respectively, and the corresponding root mean square errors are 7.66 pmol m⁻² s⁻¹, 3.40 μmol m⁻² s⁻¹, and 2.06 μmol m⁻² s⁻¹ respectively. As the time scale extends to 3 h or 6 h, the predictive performance drops significantly. At the 6-hour scale, the R² of the model for predicting COS, GPP, and NEE are only 0.23, 0.20, and 0.17 respectively. Fortunately, the predictive performance rebounds significantly at the daily scale and further strengthens at the weekly scale. At this time, the R² of the three reach 0.86, 0.85, and 0.63 respectively. Furthermore, it was found that the main factors influencing the fluxes of COS and CO₂ are completely different at different time scales. At shorter time scales, photosynthetically active radiation dominates, while at longer time scales, soil temperature and moisture become more critical influencing factors. These findings provide important clues for mechanistically simulating carbon fluxes in boreal forests.