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

Abstract

Ammonia (NH₃) is a gaseous pollutant with significant environmental and health implications. Over recent decades, its increasing concentrations, driven by industrialization and agriculture, have necessitated high-resolution monitoring. However, limited daily ground-based observations hinder comprehensive analysis. This study developed machine learning-based frameworks—deep neural network (DNN), random forest, and light gradient boosting machine—to predict biweekly NH₃ concentrations and downscale them to daily estimates across the United States during 2017–2022. The models integrate NH₃ column concentrations, meteorological variables, land cover data, livestock information, and ground-based measurements. Among the models, DNN showed superior performance in both spatial cross-validation and independent testing, achieving a correlation coefficient of 0.79, a root mean square error of 0.98 µg/m³, and an index of agreement of 0.83— effectively capturing fine-scale spatial variations at a 9 km resolution. Shapley additive explanations analysis identified temporal dynamic factors—such as day of year and meteorological variables—as key predictors, along with land cover and cattle density, highlighting the model’s ability to support the temporal downscaling of NH₃ from biweekly to daily scale. When applied to the UK, the model demonstrated its potential for spatial transferability via the leave-one station-out and leave-one year-out cross validations. These findings highlight the ability of machine learning in bridging temporal gaps and generating high-resolution daily NH₃ estimates.