Agriculture and Food Systems - Theses
Permanent URI for this collection
Search Results
1 - 2 of 2
-
ItemUsing data mining to improve the prediction of key nitrogen loss from terrestrial ecosystems( 2021)Nitrogen (N) losses through nitrification, denitrification and nitrate leaching in terrestrial ecosystems reduce fertiliser N use efficiency (NUE), affect environmental quality and human health. In the past 50 years, many individual studies both on-site and in laboratories have been carried out to investigate N loss from different pathways. However, field measurements of N dynamics are generally time-consuming, expensive and some of the processes are quite difficult to measure. Process-based models have been developed to better understand the complex, multivariate and unpredictable N cycling by integrating soil, environmental and management factors. However, the performances of these models are restricted by data availability, inconsistent responses of N loss to key drivers, difficulty of parameter derivation and limited capacity of larger scale simulation. To address the limitations of process-based models, improve the prediction and simulation of N losses, increase NUE and environmental quality, comprehensive databases of N loss from different pathways were compiled by data mining, advanced machine learning models were developed to reflect the linkage between N loss pathways and soil, environmental and climatic conditions on global scale. The research reported in this thesis quantified the nitrification rate (Rnit) and the fraction of nitrous oxide (N2O) from nitrification (fN2O_nit), denitrification and associated N2O, dinitrogen (N2) production, the contribution of N2O from autotrophic nitrification, heterotrophic nitrification and denitrification and nitrate (NO3-) leaching, investigated the key drivers of N loss from each pathway, performed global performed accurate global prediction of nitrification rate and the fraction of N2O from nitrification, NO3-N leaching with fewer input variables using data mining, machine learning models and 15N tracing experiment. Findings include: Data mining and machine learning were integrated to predict R_nit and fN2O_nit. According to the compiled global database on Rnit and fN2O_nit, the average potential Rnit in the topsoil was 1.4 kg N ha-1 d-1, and fN2O_nit was from 0.004 to 9.19% (average 0.46%). The machine-learning based stochastic gradient boosting (SGB) model outperformed three widely used process-based models in estimating R_nit and N2O emission from nitrification by using the same input variables. SGB technique was then applied for global prediction of Rnit and fN2O_nit with only a few input variables (R2 = 0.76 and 0.55, respectively). The potential Rnit was driven by long-term mean annual temperature, soil C/N ratio and soil pH, whereas fN2O_nit by mean annual precipitation, soil clay content, soil pH, soil total N. The global fN2O_nit varied by over 200 times (0.006-1.2%), it should be adjusted according to edaphic and environmental conditions when used in process-based models or global climate models in projecting N2O emissions. A global assessment of soil denitrification rate, N2O/(N2O+N2), and their driving factors and mitigation strategies was conducted based on 225 studies (3367 observations). N loss through denitrification varied greatly across land uses and climatic regions with an average of 0.25 kg N ha-1d-1. The average emission factor of denitrification (EFD) was 4.8%. The wide range of N2O/(N2O+N2) (mean: 0.33) demonstrated that the adoption of a fixed ratio in some process-based models for estimating N2 emissions from denitrification is not suitable. N2 loss accounted for 67% of total denitrification. N loss as N2, although harmless to the environment, deserves more attention from the perspective of improving NUE. Soil denitrification rate was significantly related to soil WFPS, NO3- content and soil temperature and soil oxygen (O2) content. N2 emissions were significantly correlated with latitude, WFPS, soil mineral N and soil oxygen content. Soil oxygen content, NO3- content, organic C, C/N ratio and WFPS were the key drivers of N2O/(N2O+N2) ratio. The meta-analysis showed that optimizing N application rates, using ammonium-based fertilizers compared to nitrate-based fertilizers, biochar amendment and application of nitrification inhibitors could effectively reduce soil denitrification rate and associated N2 emissions by 34-219% and 15-226%, respectively. These findings highlight that N loss via soil denitrification and N2 emissions cannot be neglected, and that mitigation strategies should be adopted to reduce N loss and improve N use efficiency. Our study provides a solid foundation to large-scale estimations of denitrification and the refinement of relevant parameters used in the submodels of denitrification in process-based models. The contribution of N2O production pathways and its driving factors in forest soils were investigated by global data analysis and an incubation experiment with 15N tracing technique in both Australia and worldwide. Based on 13 forest soils sampled within Australia, forest soils in temperate areas had the highest N2O emission rate (19.5 ug N kg-1 soil d-1), followed by subtropical and arid soils (3.84 and 3.80 ug N kg-1 soil d-1, respectively). Heterotrophic nitrification dominated N2O production in Australian forest soils; its contribution followed the order of arid (78%) > subtropical (69%) > temperate (59%). N2O from heterotrophic nitrification was negatively related to MAT and the contribution of heterotrophic nitrification to N2O production was negatively related to soil TN and TC. These results partially agreed with the global literature data synthesis, which showed that in addition to heterotrophic nitrification (42%), denitrification (43.5%) was also a key pathway of N2O production in global forest soils. Globally, soil pH, moisture content, total N content, total C content and MAT contributed to heterotrophic nitrification and denitrification to N2O production. A machine learning model (NLNO3 model) was developed based on a global literature-based database of NO3-N leaching from field experiments (1818 observations) to predict NO3-N leaching from agroecosystems. The NLNO3 model can reliably predict NO3-N leaching using a few easily accessible input variables (R2=0.75). According to the model estimation of NO3-N leaching, the global spatiotemporal pattern and hotspots were identified. The total NO3-N leaching in agroecosystems increased from 23.2 Tg N yr-1 in 1961 to 32.8 Tg N yr-1 in 2000 and 39.7 Tg N yr-1 in 2010. Hotspots of NO3-N leaching in agroecosystems expanded from Europe in 1961 to China, South Asia and Brazil in 2000 and 2010. The high spatiotemporal heterogeneity of NO3-N leaching was mainly driven by soil properties (soil TN, soil pH, soil texture), aridity index and farming practices (N fertilization and irrigation). Results of the present research demonstrate that the capacities of data mining in better understanding the complex N cycling in terrestrial ecosystems and informing potential mitigation strategies to reduce N loss. Data mining coupled with advanced machine learning methods could not only address the limitations of process-based models and improve model simulation performances, but also provide an alternative approach in predicting N dynamics at a larger scale.
-
ItemFinding the hidden smoke: Exploring the use of digital technologies for assessing grapevine smoke contamination and taint in grapes and wine( 2021)Grapevine smoke contamination and the subsequent development of smoke taint in wine has resulted in significant financial losses for winemakers throughout the world. Unfortunately, the incidence of grapevine smoke exposure is expected to rise as the number and intensity of wildfires increase due to the effects of climate change. Wines produced from smoke affected grapes are characterised by unpleasant smoky aromas, rendering them unpalatable and therefore unprofitable. Traditionally, chromatographic techniques such as gas chromatography-mass spectrometry (GC-MS) and high-performance liquid chromatography (HPLC) have been used for assessing the levels of smoke-derived volatile phenols and their glycoconjugates in grapes and wine. However, these methods are time consuming, expensive and require destructive sample preparation as well as the use of trained personnel. Furthermore, sensory evaluation of wine samples using human panels may be subject to bias due to individual variability of the participants, as well as being expensive and time consuming as large groups of participants must be recruited and trained. In addition to this, a number of methods have been identified for ameliorating smoke taint in wine such as the use of activated carbon and reverse osmosis. While effective at reducing levels of volatile phenols for smoke taint amelioration, they are unable to act on glycoconjugates, and therefore a gradual resurgence of smoky aromas may arise as these glycoconjugates are hydrolysed back into their free active forms over time. This research therefore investigated alternative methods for assessing the degree of grapevine smoke exposure and the level of smoke taint in wine using digital technologies coupled with machine learning (ML) modelling based on artificial neural networks (ANN), and whether the use of a cleaving enzyme prior to the addition of activated carbon could be effective at ameliorating smoke taint in wine. Near-infrared (NIR) spectroscopy was used to obtain a chemical fingerprint of grape berries, leaves, must and wine. These readings were then used as inputs to develop ANN models that showed high accuracy in the classification of berries and leaves according to the level of smoke exposure and degree of taint (97% – 98%), as well as predicting the levels of smoke-derived volatile phenols and their glycoconjugates in grapes, must and wine (R = 0.98 – 0.99). Additionally, models predicting consumer responses towards smoke tainted wines using NIR berry and wine spectral readings were created which displayed high accuracy in their predictive abilities (R = 0.97 – 0.98). The results demonstrated that NIR spectroscopy coupled with ML modelling can provide accurate, rapid and non-destructive tools for assessing grapevine smoke contamination and smoke taint in wine, in addition to predicting the sensory responses of consumers towards smoke tainted wines. Furthermore, the models developed can be used together to form an integrated smoke taint detection system that growers and winemakers can use in-field or in the winery to assess grapes and wine. A low-cost electronic nose (E-nose) was used to assess the aroma potential of smoke-tainted wines. Readings from the e-nose were used as inputs to develop ML models that showed high accuracy in predicting the levels of eight volatile aromatic compounds in wine (R = 0.99), the degree of smoke aroma intensity (R = 0.97). These two models may be used together with previously developed models that predict the levels of smoke-derived volatile phenols and their glycoconjugates and 12 wine descriptors to provide winemakers with a greater picture of the degree of smoke taint and the aroma profiles of smoke-tainted wines. In addition to this, the use of a cleaving enzyme (ZIMAROM, Enologica Vason) prior to treatment with activated carbon was found to be effective in ameliorating smoke taint and may help delay the resurgence of smoky aromas by hydrolysing glycoconjugates into their free volatile phenol forms which can then be removed by the addition of activated carbon. An ANN model displaying high accuracy (98%) was also developed using the readings from the e-nose to classify wine samples according to the type of smoke-taint amelioration treatment applied to assess their effectiveness. The model may offer winemakers a cost-effective, non-destructive, rapid, and accurate tool to assess the effectiveness of smoke taint amelioration treatment by activated carbon with/without the addition of a cleaving enzyme.