Modeling enteric methane emission from dairy cows using deep learning approach

Abstract

This study explores the application of deep learning (DL) models to predict methane (CH₄) emissions from enteric fermentation in dairy cows using performance, feeding, behavioral and weather data from automated milking and feeding systems, behavioral sensors, and a public weather database. Individual CH₄ emissions were recorded using sniffer technology for up to 52 cows from October 2022 to December 2023. Long Short-Term Memory (LSTM) outperformed Convolutional Neural Network (CNN) and hybrid CNN-LSTM models when all features were available (scenario S1), achieving an R² of 0.88 and a mean bias error (MBE) of 13.55 ppm. We further tested the performance of DL models under different data availability scenarios, classifying features as "rare", "moderate", or "public" based on the effort required to obtain them. Scenario S2 excluded rare features and represented a farm with only moderate and public data. Scenario S3 included only public data. Scenario S4 extended scenario S2 by including important rare features identified through feature importance analysis. Using moderate and public data yielded reasonable model performance (R² = 0.45, MBE = 17.60 ppm). Further reducing data availability to only public data substantially decreased performance. However, when three rarely available feed-related features, i.e., feed efficiency, concentrate intake, and total dry matter intake were added to the moderate and public dataset, model performance improved substantially (R² = 0.74, MBE = 14.36 ppm). A random forest feature importance analysis confirmed the critical role of feed-related data. This study highlights the potential of DL models to predict CH₄ emissions using widely available data supplemented by a few rare ones.