Combining short-term breath measurements to develop methane prediction equations from cow milk mid-infrared spectra

•Predicting methane production based on milk mid−infrared spectra.•Calibrating various combinations of short-term breath data to optimize prediction.•Correcting methane emissions for diurnal variations improved prediction ability.•Calibrating equations on heterogeneous data is essential for large−scale prediction.•Implementing a selection to reduce methane emissions from the predictions. Predicting methane (CH4) emission from milk mid-infrared (MIR) spectra provides large amounts of data which is necessary for genomic selection. Recent prediction equations were developed using the GreenFeed system, which required averaging multiple CH4 measurements to obtain an accurate estimate, resulting in large data loss when animals unfrequently visit the GreenFeed. This study aimed to determine if calibrating equations on CH4 emissions corrected for diurnal variations or modeled throughout lactation would improve the accuracy of the predictions by reducing data loss compared with standard averaging methods used with GreenFeed data. The calibration dataset included 1 822 spectra from 235 cows (Holstein, Montbéliarde, and Abondance), and the validation dataset included 104 spectra from 46 (Holstein and Montbéliarde). The predictive ability of the equations calibrated on MIR spectra only was low to moderate (R2v = 0.22–0.36, RMSE = 57–70 g/d). Equations using CH4 averages that had been pre−corrected for diurnal variations tended to perform better, especially with respect to the error of prediction. Furthermore, pre−correcting CH4 values allowed to use all the data available without requiring a minimum number of spot measures at the GreenFeed device for calculating averages. This study provides advice for developing new prediction equations, in addition to a new set of equations based on a large and diverse population.