Note for Dutch speakers: a summary in Dutch can be downloaded here.
Production of food contributes to climate change. To develop strategies to produce food with a low environmental impact, environmental assessment models, such as life cycle assessment (LCA) or nutrient balance (NB) analysis are applied. Input data required for these models, may vary due to seasonal changes, geographical conditions or socio-economic factors. Moreover, input data may be uncertain, due to measurement errors and observational errors that exist around modeling of emissions and technical parameters. Although agricultural activities and food production are prone to natural variability and epistemic uncertainty, very few case studies made a thorough examination of the effects of variability and uncertainty on the result.
This thesis aimed to enhance understanding the effects of variability and uncertainty on the results. This was done by exploring how uncertainty analysis and sensitivity analysis can help to reduce the efforts for data collection, support the development of mitigation strategies and improve overall reliability, leading to more informed decision-making in environmental impact assessment models. To that end, methods for uncertainty analysis and sensitivity analysis were combined, and the effect of correlations in uncertainty propagation and global sensitivity analysis were explicitly accounted for. To be able to formulate case study specific suggestions that could improve reliability and point to potential mitigation strategies in food production, methods were applied to case studies of dairy and pork production.
To derive mitigation options and improve reliability in the assessment of greenhouse gas emissions (GHGs) along the pork production chain, two sensitivity methods were combined: the multiplier method and the method of elementary effects. The multiplier method showed how much the input parameter influences the assessment of GHGs, whereas the method of elementary effects showed the importance of input parameters on the output uncertainty. By combining the result of the multiplier method and the method of elementary effects, the essential parameters were identified. Results of this study showed that the most essential input parameter are the feed conversion ratio, the amount of manure, CH4 emissions from manure management and crop yields, especially maize and barley. Combining the results of both methods allowed finding mitigation options, either based on innovations (e.g. novel feeding strategies) or on management strategies (e.g. reducing mortality rate). Furthermore, reliability could be improved by increasing data quality of the most essential parameters [Groen et al., 2016].
Uncertainty propagation in environmental impact assessment models such as LCA, are usually performed using Monte Carlo sampling. However, other methods for uncertainty propagation are available, and it was unknown which method performed best. Monte Carlo sampling, Latin hypercube sampling, quasi Monte Carlo sampling, analytical uncertainty propagation and fuzzy interval arithmetic were compared based on convergence rate and output statistics. Each method was tested on three LCA case studies, which differed in size and behavior. Results showed that uncertainty propagation in LCA using a sampling method leads to more (directly) usable information compared to fuzzy interval arithmetic or analytical uncertainty propagation. Latin hypercube and quasi Monte Carlo sampling provide more accuracy in determining the population mean than Monte Carlo sampling and can even converge faster than Monte Carlo sampling for some of the case studies discussed [Groen et al., 2014]
Global sensitivity analysis in environmental impact assessment models, such as LCA, can be performed using several different methods. However, which method is most suitable was unknown. Five methods and coefficients that can be used for global sensitivity analysis were compared: standardized regression coefficient, Spearman correlation coefficient, key issue analysis, Sobol’ method and random balance design. To be able to compare the performance of global sensitivity methods, two hypothetical case studies were constructed. The comparison of the sensitivity methods was based on four aspects: (I) sampling design, (II) output variance, (III) explained variance, and (IV) contribution to output variance of individual input parameters. Key issue analysis does not make use of sampling and was fastest, whereas the Sobol’ method had to generate two sampling matrices, and therefore, was slowest (I). The total output variance (II) resulted in approximately the same output variance for each method, except for key issue analysis, which underestimated the variance especially for high input uncertainties. The explained variance (III) and contribution to variance (IV) for small input uncertainties, was optimally quantified by standardized regression coefficients and the main Sobol’ index. For large input uncertainties, Spearman correlation coefficients and the Sobol’ indices performed best. We concluded that the standardized regression coefficients, Spearman correlation coefficients or key issue analysis could be used for global sensitivity analysis in environmental impact assessment models [paper under review].
Incorporation of uncertainty propagation in LCA is nowadays widely acknowledged. Currently, most LCA studies that include uncertainty analysis ignore correlations between input parameters during uncertainty propagation, due to unfamiliarity with methods that include correlations or lack of data. The effect of ignoring these correlations on the output variance, however, remains unclear: it is not known if and under which conditions it can lead to erroneous conclusions. Two approaches to include correlations between input parameters during uncertainty propagation were studied: an analytical approach and a sampling approach. The use of both approaches is illustrated for an artificial case study of electricity production. Results demonstrated that that both approaches yield approximately the same output variance and sensitivity indices for this specific case study. Furthermore, we demonstrated that the analytical approach can be used to quantify the risk of ignoring correlations between input parameters during uncertainty propagation in LCA. We concluded that: (1) we can predict if including correlations among input parameters in uncertainty propagation will increase or decrease output variance; (2) we can quantify the risk of ignoring correlations on the output variance and the global sensitivity indices. Moreover, this procedure requires only little data regarding the input parameters [Groen and Heijungs, 2016].
LCA of dairy products such as milk, require many input parameters that are often affected by variability and uncertainty. Moreover, correlations may be present between input parameters, e.g. between feed intake and milk yield. Three diets corresponding to three grazing systems (zero-, restricted and unrestricted grazing) were selected, which were defined to aim for a given milk yield. First, a local sensitivity analysis was used to identify which parameters influence GHG emissions most. Second, a global sensitivity analysis was used to identify which parameters are most important to the output variance. The global sensitivity analysis included correlations between feed intake and milk yield and between nitrogen (N) fertilizer rates and crop yields. The local and global sensitivity analyses were combined to determine which parameters are essential. Finally, we analysed the effect of changing the most important correlation coefficient (between feed intake and milk yield) on the output variance and global sensitivity analysis. The mean GHG emissions for 1 kg energy corrected milk ranged from 1.08 to 1.12 kg CO2 equivalents, depending on the grazing system. The most essential parameters were the CH4 emission factor of enteric fermentation, milk yield, feed intake, the direct N2O emission factor of crop cultivation For both grazing systems, the N2O emission factor for grazing also turned out to be important. In addition, the correlation coefficient between feed intake and milk yield turned out to be important. Moreover, systematic combining the local and global sensitivity analysis resulted in more parameters than previously found [Wolf and Groen et al., 2016].
A nutrient balance quantifies differences in nutrients entering and leaving the system and can be expressed in e.g. nutrient use efficiency (NUE). NUE is commonly used to benchmark the environmental performance of dairy farms. Benchmarking farms, however, may lead to biased conclusions because of differences in major decisive characteristics between farms, such as soil type and production intensity, and because of epistemic uncertainty of input parameters. To compare NUE of farms, farms were clustered based on similar characteristics, which resulted in farming systems. Farming system 1 was located on sandy soils and were more intensive in terms of milk production than farms in farming system 2. Farming system 2 was located on loamy soils. First, Monte Carlo sampling was used to propagate input uncertainties through the nutrient balance. Including the epistemic uncertainty of input parameters showed that benchmarking NUE of farms in farming system 1 was no longer possible, whereas farms in farming system 2 could still be ranked when uncertainty was included. Second, a global sensitivity analysis was performed to quantify how much the input parameters contributed to the output variance, using the squared standardized regression coefficients. Input parameters that explained most of the output variance differed between framing systems. For farming system 1, input of feed and output of roughage were most important. For farming system 2, the input of mineral fertilizer was most important. Third, the uncertainty of the parameters explaining most of the output variance was reduced to examine if this would improve benchmarking results. After reducing the uncertainties of the most important parameters, benchmarking results significantly improved [paper under review].
Improving the value of uncertainty analysis and sensitivity analysis in environmental impact assessment models, specifically in LCA and NB analysis, would benefit from standardizing the use of definitions and methods. Currently in the ISO standard for LCA, there is no method recommended for neither a local nor global sensitivity analysis. Developing a standardized definition and method for global sensitivity analysis in LCA and NB studies, will increase comprehensibility between studies and will enhance the comparison of results between studies. The uncertainty analysis or sensitivity analysis to be applied depends on the question to be addressed and the available information. However, in most environmental impact assessment models, data availability is limited and combining local and global sensitivity analysis makes sure that parameters are not overlooked. A local and global sensitivity analysis should, therefore, be seen as complementary. Moreover, we concluded that both sampling approach for uncertainty propagation and analytical uncertainty propagation are indispensible in environmental impact assessment models. The analytical approach is especially useful when data is limited. In contrast, the sampling approaches are more suitable when full knowledge is available [Discussion].
This thesis showed that using a systematic approach to uncertainty and sensitivity analysis improves overall reliability, reduces efforts for improved data collection and supports the development of potential mitigation strategies, especially for case studies of food production, where epistemic uncertainty and variability are ubiquitous.
Source: Summary PhD thesis Evelyne Groen, An uncertain climate: the value of uncertainty and sensitivity analysis in environmental impact assessment of food, 2016
ISBN: 978-94-6257-755-8; DOI: 10.18174/375497