当前位置: X-MOL 学术Agron. J. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Bayesian hybrid analytics for uncertainty analysis and real-time crop management
Agronomy Journal ( IF 2.1 ) Pub Date : 2021-03-22 , DOI: 10.1002/agj2.20659
Esther D. Meenken 1 , Christopher M. Triggs 2 , Hamish E. Brown 3 , Sarah Sinton 3 , Jeremy R. Bryant 4 , Alasdair D.L. Noble 1 , Martin Espig 1 , Mostafa Sharifi 1 , David M. Wheeler 4
Affiliation  

Dynamic, deterministic agricultural models, and current machine learning technologies based on sensor data, enable and support decision making for on-farm management. However, their predictions are subject to various sources of uncertainty. Hybrid analytics that leverage both modelled and sensor data provide predictive information that makes the best of both approaches in a timely fashion to inform operational decision making and enable inclusive uncertainty quantification. We describe and evaluate a probabilistic Bayesian data assimilation tool that combines the state variables from the Sirius wheat (Triticum aestivum L.) development model with time-series environmental and leaf count data. Additionally, the uncertainty associated with input parameters is quantified via expert opinion. The Bayesian approach obtained point estimates through time that were accompanied by inclusive, probabilistic, 95% credible intervals. At the end of simulation, a typical model predicted a final leaf number of 6.6 leaves, Sirius alone predicted seven leaves and the mean of the observed data was 6.7 leaves. The 95% credible interval was estimated as 5.1–8.4 leaves. Importantly, the tool was able to “redirect” simulated outputs if input parameters such as minimum leaf number or base phyllochron were incorrectly specified, with the implication that on-farm decision makers would have advance warning of variation in expected harvest date. Relatively few plants with time-intensive data were sufficient to fit the model, however, more plants would be desirable to reduce the rather wide range of credible intervals. Nevertheless, the tool shows potential and could be readily implemented with low resource requirements, providing more finely tuned harvest date information, with probabilistic uncertainty quantification built-in, for on-farm decisions.

中文翻译:

用于不确定性分析和实时作物管理的贝叶斯混合分析

动态的、确定性的农业模型和基于传感器数据的当前机器学习技术,支持并支持农场管理的决策制定。然而,他们的预测受到各种不确定性来源的影响。利用建模数据和传感器数据的混合分析可提供预测信息,及时充分利用这两种方法,为运营决策提供信息,并实现包容性的不确定性量化。我们描述并评估了一种概率贝叶斯数据同化工具,该工具结合了来自Sirius小麦 ( Triticum aestivumL.) 具有时间序列环境和叶数数据的开发模型。此外,与输入参数相关的不确定性通过专家意见进行量化。贝叶斯方法通过时间获得点估计,并伴随有包容性、概率性、95% 的可信区间。模拟结束时,典型模型预测最终叶数为 6.6 片,天狼星单独预测 7 片叶子,观察数据的平均值为 6.7 片叶子。95% 的可信区间估计为 5.1-8.4 片叶子。重要的是,如果输入参数(例如最小叶数或基叶轮距)被错误指定,该工具能够“重定向”模拟输出,这意味着农场决策者将提前警告预期收获日期的变化。具有时间密集型数据的相对较少的工厂足以拟合模型,但是,需要更多的工厂来减少相当广泛的可信区间。尽管如此,该工具显示出潜力,并且可以在低资源需求的情况下轻松实施,提供更精细调整的收获日期信息,内置概率不确定性量化,用于农场决策。
更新日期:2021-03-22
down
wechat
bug