In this paper, we discuss the use of Automated Machine Learning for the first time applied to an Ex-ante Life Cycle Assessment. This kind of analysis has been conducted for a particular crop production, i.e. barley. The aim is to assess the impact (in terms of carbon dioxide emissions and yield) of diferent production strategies. The data used in this study comes from a two-year measurement campaign involving five countries. The results are compared against the state-of-the-art technique, showing the good performance of the approach.
Ex-ante Life Cycle Assessment (LCA), unlike traditional
LCA, is conducted before a product, process, or service
is fully developed or implemented. Hence, it refers to a
predictive or foresight-oriented approach to
environmental impact assessment. While traditional LCA assesses the
environmental impacts of an existing product or process
(post-implementation), ex-ante LCA is performed during
the design phase or early planning stages. It aims to forecast
potential environmental impacts based on available data and
assumptions. It may help decision-makers evaluate diferent
alternatives and make informed choices about which design
or process will likely have the least environmental impact.
Since ex-ante LCA is conducted early in the product
lifecycle, it often relies on estimated data or assumptions that
might not be as accurate as data obtained after production
or use. This introduces some uncertainty in the analysis, but
it still provides valuable insights for comparison. It allows
for the early identification of potentially unsustainable
practices, guiding toward more sustainable solutions. It brings
some benefits, since assessing environmental impacts early
in the process can help avoid costly mistakes and optimize
the environmental performance of a product or service. The
limitations mainly deal with uncertainty, since early-stage
assessments are based on models and assumptions that may
not fully reflect the actual environmental impact once the
product or process is realized [
In this work, we explore the integration of an Automated
Machine Learning (AutoML) technique to enhance ex-ante
LCA, specifically for predicting crop yields and CO2
emissions. AutoML provides an eficient framework for
automating the process of model selection and hyperparameter
optimization, enabling rapid development of predictive models
with minimal manual intervention [
The remainder of this paper is organized as follows. Section 3 introduces the methods and the data collection. Section 4 describes the details of our experimental setup and the evaluation of the proposed methodology, before concluding the paper in Section 5.
By looking at the definitions in the literature, the ex-ante
LCA aims to [
Similarly, it can be regarded as an “environmental [LCA]
of a new technology before it is commercially implemented
in order to guide R& D decisions to make this new
technology environmentally competitive as compared to the
incumbent technology mix” [
The challenging part of conducting an ex-ante LCA is the
lack of representative data for the system assessed which
could bring uncertainty to the study [
Data comes from a two-year measurement campaign, over
2022 through 2023, in the context of the SusCrop
ERANET funded project ConnectFarms [
There were two fertilizer treatments: non-fertilized treatment (control) and mineral fertilized treatment N23P30K63. All treatments were in four replications. The plot size was 20 m2. Spring barley yield was 3,8 t/ha on average in fertilized treatment, which was significantly higher than in non-fertilized control (Figure 2). The highest yields were in Turkey (4,2 t/ha), being the south-most testing site. Most likely irrigation helped to achieve the result. Interestingly, fertilization did not afect the yield. In Bulgaria, there were high yields in fertilized treatments (average 3,8 t/ha), which was almost twice as high as controls. Twice the diference in yields was also observed in Estonia.
In order to build our ML model, the considered variables and their range are shown in Table 1. The yield and carbon dioxide emissions represent the output of our model. The whole dataset, in the ranges shown, has 2,000 samples.
The data was normalized for each variable as = ( − )/, where is the mean and is the standard deviation. Name
Seeds Harrowing Harvesting
Ploughing Sowing or planting
Nitrogen Phosphate Potassium Compost Biochar Yield CO2
Unit
kg
h
h
h
h
g/m2
g/m2
g/m2
kg/m2
kg/m2
kg/m2
ppm
We leverage Auto-sklearn [
The state-of-the-art baseline model is the Co-Active
NeuroFuzzy Inference System with fractional regularization
(CANFIS-T) proposed in [
Hardware Resources. We conducted our experiments on a CPU environment. The CPU environment runs on Windows 11 Pro 64-bit (10.0, Build 22621) with 16 core Intel(R) Core(TM) i9-10885H Processor @ 2.40GHz,32 GB DIMM memory, and 1000 GB SSD data storage. All the approaches have been implemented in Python.
Treatment Type Biochar Treatment
NPK Treatment No Fertilizer Treatment
MSE (train) -11.9693 -0.0073 -0.0041
RMSE (train) 3.4586 0.0848 0.0627
MSE (test) 10.8 753 0.0053 0.0029
RMSE (test) 3.2978 0.0728 0.0542 2 Score (test) 0.9589 1.0000 1.0000 The results of the AutoML models applied for predicting CO2 emissions reveal significant variances in performance across diferent treatment types, as shown in Table 2. The Biochar Treatment Model achieved a mean R² score of 0.9589, indicating that it explains approximately 95.9% of the test set’s variance. Despite this strong correlation, the model demonstrated a root mean square error (RMSE) of 3.2978, suggesting that while the model is reliable, there may be underlying complexities in the biochar data or the target variable distribution contributing to the prediction errors. In contrast, both the NPK Treatment Model and the No Fertilizer Treatment Model reached a perfect R² score of 1.0000, with RMSE values of 0.0728 and 0.0542, respectively. These results imply that the models accurately predict CO2 emissions in these scenarios without any observable error, highlighting their robustness. The exceptional performance of the NPK and No Fertilizer models suggests that these treatment types exhibit more consistent patterns in CO2 emissions compared to the biochar treatment. Overall, these ifndings show the efectiveness of the AutoML approach in modelling carbon emissions under varying agricultural practices, with implications for optimizing treatments aimed at reducing greenhouse gas emissions.
The results of the yield prediction models indicate varying levels of performance across the diferent treatment types, as shown in Table 3. The Biochar model produced a mean R² score of 0.8956, reflecting that approximately 89.6% of the variance in the test set can be explained by the model. This model exhibited a root mean square error (RMSE) of 0.2331, suggesting some level of prediction error, despite the robust performance indicated by its mean squared error (MSE) of 0.0543. In contrast, the NPK model demonstrated a higher R² score of 0.9481, meaning it accounts for about 94.8% of the variance in yield, with an RMSE of 0.4674, indicating slightly larger errors in predictions compared to the Biochar model. This model’s mean MSE of 0.2185 is also higher, suggesting that while it performs well, it may be less precise than the Biochar treatment. The No Fertilizer model showed the lowest performance among the three, with an R² score of 0.8593, indicating it explains approximately 85.9% of the variance in the test set. The RMSE of 0.3063 indicates notable prediction errors. Collectively, these results highlight the efectiveness of the AutoML approach in predicting agricultural yield, with the NPK treatment showing the highest explanatory power, followed by Biochar and No Fertilizer treatments.
The study included a comparison against the CANFIS-T
method, previously proposed to tackle the same problem in
[
When compared with the baseline CANFIS-T method, the AutoML framework substantially outperforms it in terms of predictive accuracy. For example, the AutoML models yielded lower RMSE values (0.2331 for Biochar and 0.4674 for NPK) compared to the CANFIS-T method’s results, such as 0.721643 for Biochar and 2.1265502 for NPK. These discrepancies highlight the superior generalization ability of AutoML in accurately modelling agricultural data across diferent treatment types. Additionally, while the CANFIS-T method demonstrated faster training times, its performance in CO2 emission prediction was less favourable, with an RMSE of approximately 10 across all cases. This further underscores the eficacy of AutoML in addressing complex prediction tasks while maintaining a high level of model stability and accuracy, particularly in cases with higher data variability.
Overall, the findings suggest that AutoML ofers significant advantages in both predictive performance and model generalization over traditional methods, providing valuable insights for optimizing agricultural practices aimed at mitigating greenhouse gas emissions.
In future work, we will investigate the use of AutoML techniques to handle uncertainty and noisy data in agriculture.
This work has been partially funded by the Estonian Research Council, grant PRG1604, through the funding of SusAn, FACCE ERA-GAS, ICT-AGRI-FOOD and SusCrop ERA-NET, and through the project Increasing the knowledge intensity of Ida-Viru entrepreneurship co-funded by the European Union.