Comparing different statistical models for predicting greenhouse gas emissions, energy-, and nitrogen intensity

Summary in the original language of the document

To evaluate the environmental impact across multiple dairy farms cost-effectively, the methodological framework for environmental assessments may be redefined. This article aims to assess the ability of various statistical tools to predict impact assessment made from a Life Cyle Assessment (LCA). The different models predicted estimates of Greenhouse Gas (GHG) emissions, Energy (E) and Nitrogen (N) intensity. The functional unit in the study was defined as 2.78 MJMM human-edible energy from milk and meat. This amount is equivalent to the edible energy in one kg of energy-corrected milk but includes energy from milk and meat. The GHG emissions (GWP100) were calculated as kg CO2-eq per number of FU delivered, E intensity as fossil and renewable energy used divided by number of FU delivered, and N intensity as kg N imported and produced divided by kg N delivered in milk or meat (kg N/kg N). These predictions were based on 24 independent variables describing farm characteristics, management, use of external inputs, and dairy herd characteristics.
All models were able to moderately estimate the results from the LCA calculations. However, their precision was low. Artificial Neural Network (ANN) was best for predicting GHG emissions on the test dataset, (RMSE = 0.50, R2 = 0.86), followed by Multiple Linear Regression (MLR) (RMSE = 0.68, R2 = 0.74). For E intensity, the Supported Vector Machine (SVM) model was performing best, (RMSE = 0.68, R2 = 0.73), followed by ANN (RMSE = 0.55, R2 = 0.71,) and Gradient Boosting Machine (GBM) (RMSE = 0.55, R2 = 0.71). For N intensity predictions the Multiple Linear Regression (MLR) (RMSE = 0.36, R2 = 0.89) and Lasso regression (RMSE = 0.36, R2 = 0.88), followed by the ANN (RMSE = 0.41, R2 = 0.86,). In this study, machine learning provided some benefits in prediction of GHG emission, over simpler models like Multiple Linear Regressions with backward selection. This benefit was limited for N and E intensity. The precision of predictions improved most when including the variables “fertiliser import nitrogen” (kg N/ha) and “proportion of milking cows” (number of dairy cows/number of all cattle) for predicting GHG emission across the different models. The inclusion of “fertiliser import nitrogen” was also important across the different models and prediction of E and N intensity.