Data-driven approach to estimate ambient ammonia concentrations

The Netherlands experiences significant environmental impacts from reactive nitrogen, particularly ammonia (NH 3 ), whose long-term deposition harms soil, water, and biodiversity. NH 3 is emitted mainly from agriculture and its levels can be very high and variable in intensive livestock areas. Although monitoring stations are set up across multiple locations, data gaps may still arise due to issues with measurement instruments. Therefore, there is a need for models that can reliably fill in these missing observations. In this study, we first applied Linear Regression, Random Forest, and Gated Recurrent Unit (GRU) models to predict one day ahead NH 3 concentrations at six sites (De Zilk, Valthermond, Vredepeel, Wekerom, Wieringerwerf, Zegveld) in the Netherlands. All models captured daily NH 3 trends, with Random Forest initially performing best (R 2 up to 0.59 at Valthermond). Predictions were strongly driven by lagged NH 3 values with wind speed and precipitation contributing moderately. After correcting for prediction lag caused by autocorrelation, the linear regression performed best at most sites (with best performance observed at Zegveld with R 2 of 0.84) and was used for gap filling. It maintained reasonable accuracy despite sparse observations (every 2–7 days), with most sites (De Zilk, Valthermond, Vredepeel, Wekerom, Zegveld) exhibiting a significant correlation of almost or more than 0.5 even at 7 days measurement gaps. Additionally, we noted that historical data from existing sites could be effectively used to fill gaps at newly established locations that exhibited similar distributions of NH 3 .