Comparative Prediction of Methane Production In Vitro Using Multiple Regression Model and Backpropagation Neural...

Area of Science:

• Animal Science
• Agricultural Engineering
• Environmental Science

Background:

• Methane (CH4) production from rumen fermentation contributes to global warming and represents energy loss in beef cattle.
• Accurate prediction of CH4 emissions is essential for mitigating environmental impact and improving feed efficiency.

Purpose of the Study:

• To compare the predictive accuracy of multiple linear regression (MLR) and backpropagation neural network (BPNN) models for ruminal CH4 production.
• To evaluate these models based on carbohydrate (Carbs) components within the Cornell Net Carbohydrate and Protein System (CNCPS) framework.
• To assess model performance across various concentrate-to-forage (C/F) ratios in beef cattle rations.

Main Methods:

• Two datasets were generated using the Menke and Steingass in vitro fermentation method.
• One dataset (60 rations) was used for model development across diverse C/F ratios (30:70 to 90:10).
• A separate dataset (10 rations) was used for model validation and comparative accuracy assessment.

Main Results:

• Both MLR and BPNN models showed significant relationships between CH4 production and CNCPS carbohydrate components (CA, CB1, CB2, CC).
• MLR model achieved an R-squared of 0.91 (p < 0.0001).
• An optimal BPNN model (2 hidden-layer neurons) yielded a higher R-squared of 0.93 (p < 0.0001), indicating superior prediction performance with a lower RMSPE and higher CCC.

Conclusions:

• Both MLR and BPNN models are suitable for predicting CH4 production using CNCPS carbohydrate components.
• The BPNN model demonstrated superior predictive accuracy compared to the MLR model.
• These findings support the use of advanced modeling techniques like BPNN for more precise CH4 emission estimations in cattle.