Bridging temporal gaps: AI-based temporal downscaling of biweekly NH3 to daily scale with spatial transferability

Ammonia (NH₃) is an important atmospheric pollutant with environmental and public health impacts. In recent decades, NH₃ concentrations have increased due to intensified agricultural activities and industrial development, underscoring the need for high-resolution monitoring. However, sparse biweekly ground-based observations from the Ammonia Monitoring Network (AMoN) remain a major limitation for comprehensive spatiotemporal analysis. The United States (US) is a region where NH₃ monitoring is particularly important due to its extensive agricultural activities. In this study, we developed machine learning–based frameworks, including a deep neural network (DNN), random forest, and light gradient boosting machine, to estimate nationwide biweekly NH₃ concentrations and temporally downscale them to daily values across the contiguous US from 2017 to 2022. The models incorporate satellite-derived NH₃ column measurements, meteorological variables, land cover characteristics, livestock density, and AMoN ground-based observations. Among the tested approaches, the DNN demonstrated the strongest performance under both spatial cross-validation and independent testing, achieving a correlation coefficient of 0.79, a root mean square error of 0.98 µg m⁻³, and an index of agreement of 0.83. The model effectively reproduced fine-scale spatial variability in daily NH₃ concentrations at a 9 km resolution. Shapley additive explanations further revealed that temporally varying predictors—such as day of year and meteorological conditions—played a dominant role, alongside land cover and cattle density, supporting robust temporal downscaling from biweekly to daily scales. To assess spatial transferability, the framework was additionally applied to ground-based monitoring stations in the United Kingdom, where daily NH₃ observations are available, using leave-one-station-out and leave-one-year-out cross-validation schemes. Overall, our results demonstrate the potential of machine learning approaches to bridge temporal gaps in NH₃ observations and to generate high-resolution daily concentration estimates.