Predicting dry matter intake in gestating ewes using greenhouse gas measurements from portable accumulation chambers

Accurate measurement of dry matter intake (DMI) in sheep is logistically challenging and costly, particularly under commercial conditions and across diverse production systems. This study evaluated whether methane (CH4), carbon dioxide (CO2) and oxygen (O2) gas measurements from portable accumulation chambers (PAC), combined with bodyweight (BW) and eating time (ET), can predict DMI in gestating ewes. Two indoor feeding trials were conducted at a single research facility using Norwegian White Sheep (NWS) and Old Norwegian Spaelsau (ONS). Trial 1 used two grass silage qualities in a 2 × 2 factorial design; Trial 2 used fresh-cut herbage in a repeated-measures design. Ewes (n = 40 per trial; 29 in common) were individually housed with access to one feed bin and fed ad libitum. Daily DMI was recorded using automated feed bins, and PAC measurements of CH4, CO2, and O2 were collected on 20 d per ewe per trial. After quality control, 1,830 ewe d records remained. Supervised learning models were trained on ewe-identity-blocked training sets and evaluated on external test sets stratified by trial and breed. In this particular dataset, random forest (RF) models outperformed regularized regression, partial least squares, k-nearest neighbors, support vector machines, and a deep neural network. The best RF model, using PAC traits, LW, and ET (Pred_DMI3), achieved R2 = 0.77, root mean square error of 279 g/d, and mean absolute percentage error = 13.8% in the combined test set. Models using only PAC traits performed comparably to models including BW and ET with only marginal losses in validation metrics. Mixed-model analyses indicated strong individual-level correlations between predicted and observed DMI (rᵢ = 0.79-0.81) after adjusting for breed, diet, and trial. These results demonstrate that short-term PAC-derived gas traits can provide scalable proxies for daily DMI in gestating ewes under controlled indoor conditions, with potential application in intake phenotyping and benchmarking.