Dynamic selection of environmental variables to improve the prediction of aphid phenology: A machine learning approach
Dynamic selection of environmental variables to improve the prediction of aphid phenology: A machine learning approach
复制标题
DOI：
10.1016/j.ecolind.2017.10.032
复制
发表时间：
2018-05-01
影响因子：
6.9
通讯作者：
Bell, James R.
中科院分区：
文献类型：
--
作者：
Holloway, Paul;Kudenko, Daniel;Bell, James R.
鉴于最近杀虫剂耐药性的出现、有效化合物（如新烟碱类）从市场上消失以及气候变化促进害虫成功越冬和提前迁徙，害虫现在对作物生产构成了更大的威胁。作为监测工具，预测模型对于缓解虫害暴发具有重要意义。目前，它们为物种的出现、分布和迁移模式提供决策支持，它们的有效使用使种植者有更多时间采取战略性作物干预措施，如延迟播种或有针对性地使用杀虫剂。现有技术可能已经达到了其最佳用途，特别是在复杂系统和不断变化的气候中。机器学习（ML）可以说是对现有能力的一种进步，因为它有可能有效地识别最具信息量的时间窗口，同时改进物种预测。这样做，当种植者可以干预时，ML可能会提前任何综合虫害管理机会的长度。作为一个例子，我们研究了51种蚜虫的迁移，其中包括世界上一些最重要的经济害虫。我们使用熵和C5.0增强决策树的组合来确定最具信息量的时间窗口，将气象变量与英国各地的蚜虫迁移模式联系起来。与一般加性模型相比，决策树将首飞预测精度显著提高了20%；此外，与专家导出的变量相比，由熵选择的气象变量显著提高了3-5%的准确性。较粗的（如月）天气变量与较细的（如日）变量的精度相似，但最准确的模式包括不同周期长度的多个时间分辨率。这种综合分辨率模型本身就突出了机器学习准确预测物种及其气象驱动因素之间复杂关系的能力，这在很大程度上超出了该领域专家的经验。最后，我们确定了这些模型预测长期首飞模式的潜力，其中机器学习获得了与短期预测同等高的预测能力。虽然机器学习是统计学上的进步，但它不一定是万灵药：专家将需要用机械的理解来支撑结果，从而避免虚假的关系。本研究结果为研究人员在预测生态现象时推导和选择最合适的环境变量提供了一种自动化方法，同时提高了模型的准确性。
Insect pests now pose a greater threat to crop production given the recent emergence of insecticide resistance, the removal of effective compounds from the market (e.g. neonicotinoids) and the changing climate that promotes successful overwintering and earlier migration of pests. As surveillance tools, predictive models are important to mitigate against pest outbreaks. Currently they provide decision support on species emergence, distribution, and migration patterns and their use effectively gives growers more time to take strategic crop interventions such as delayed sowing or targeted insecticide use. Existing techniques may have met their optimal usefulness, particularly in complex systems and changing climates. Machine learning (ML) arguably is an advance over current capabilities because it has the potential to efficiently identify the most informative time windows whilst simultaneously improving species predictions. In doing so, ML is likely to advance the length of any integrated pest management opportunity when growers can intervene. As an example, we studied the migration of 51 species of aphids, which include some of the most economically important pests worldwide. We used a combination of entropy and C5.0 boosted decision trees to identify the most informative time windows to link meteorological variables to aphid migration patterns across the UK. Decision trees significantly improved the accuracy of first flight prediction by 20% compared" to general additive models; further, meteorological variables 'that were selected by entropy significantly improved the accuracy by a further 3-5% compared to expert derived variables. Coarser (e.g. monthly) weather variables resulted in similar accuracies to finer (e.g. daily) variables but the most accurate model included multiple temporal resolutions with different period lengths. This combined resolution model alone highlights the ability of machine learning to accurately predict complex relationships between species and their meteorological drivers, largely beyond the experience of experts in the field. Finally, we identified the potential of these models to predict long-term first flight patterns in which machine learning attained equally high predictive ability as shorter-term forecasts. Whilst machine learning is a statistical advance, it is not necessarily a panacea: experts will be needed to underpin results with a mechanistic understanding, thus avoiding spurious relationships. The results of this study should provide researchers with an automated methodology to derive and select the most appropriate environmental variables when predicting ecological phenomena, while simultaneously improving the accuracy of such models.