Abstract: Returning manure to the field is one of the most crucial links to integrate livestock and crop farming in the resource recycling of precise agriculture. However, its ammonia (NH₃) emission has posed a serious threat to the ecological environment. It is often required to reduce nitrogen use efficiency and environmental pollution caused by ammonia emission during manure application. It is also urgent to define the low-ammonia-emission boundaries for the return of manure to the field. The effective supply of nitrogen to crops can greatly contribute to the optimal measure to minimize the ammonia emissions. This study aims to develop a technology chain with the "low-ammonia-emission boundary definition—emission factor prediction—optimization strategy generation" using machine learning. Ammonia emissions were also reduced to optimize the practices for the effective nitrogen supply of the crops. Specifically, the emission factors (E_EF) and cumulative emissions (E_CE) were used as the input features to cluster emission data. The samples were then classified into high- and low-emission groups. Subsequently, Support vector regression (SVR) was employed to define the low-emission boundaries. A systematic investigation was made to explore the nonlinear relationships between variables. Additionally, a prediction model was developed for the ammonia emission factors from manure application. A multi-head attention mechanism was also integrated into the neural network to predict the ammonia emission factors. Finally, a low-ammonia-emission optimization was established, according to the prediction model and low-emission boundaries. The key parameters were adjusted to reduce the ammonia emissions, such as manure application, dry matter content, and pH value. The results showed that the boundary range was defined between 0.160 and 0.170 for the low emission factors (E_EF), which was lower than both the default value of 0.210 for the livestock manure emission factors specified in the IPCC guidelines and the average of 0.190 reported in previous studies. A more precise low-emission target was then provided for the agricultural manure returning to the field. In the high-emission group, E_EF values were concentrated above 0.150, and most E_CE values exceeded 50 kg/hm². In contrast, E_EF values of the low-emission group were distributed mainly between 0.050 and 0.150, whereas E_CE values were concentrated between 10 and 20 kg/hm², indicating the significant differences between the two groups. The prediction model of the ammonia emission factor with the neural network also exhibited high robustness to predict the low-emission factor region, where the data points were generally distributed around the ideal fitting line. The optimization strategy generator (that constructed using the low-emission boundaries and prediction model) was prioritized adjustments to the manure application (from surface application to band application, which was improved ammonia emission reduction efficiency by 50%-80%), followed by adjustments to dry matter content (e.g., reducing from 0.320% to 0), and finally adjustments to pH value (e.g., decreasing from 7.7 to 6.2; a pH value of 5.5 was maintained to significantly reduce the ammonia emissions). Notably, the emission factor was particularly reduced in the samples with the medium to high nitrogen application rates (150-250 kg/hm²), leading to a substantial decrease in the ammonia emission factors. In conclusion, the optimization strategy generator on the low-emission boundaries and the prediction model can be expected to optimize the manure application and properties using feature adjustments, significantly reducing the ammonia emission factors. This finding can also provide technical support for the precision fertilization and ammonia emission reduction.