Using Artificial Intelligence to Promote Food Security

在人口不断增长和气候不断变化的情况下，维持粮食安全是人类面临的最重要挑战之一。在农业方面，这要求我们在增加作物产量的同时减少化学品投入。精准农业是解决这一问题的最有希望和令人兴奋的方法之一，它允许有针对性地应用化学品，以尽量减少对环境的影响。这个博士项目应用尖端的人工智能方法来了解植物营养和胁迫，使用曼彻斯特开发的一种新型传感器技术-正弦调制荧光成像（SMFI）。SMFI在一项研究中为每一种植物生成大数据集（t字节），该数据集由一系列复杂的（相量）谐波参数组成，用于成像作物样本中每个荧光激发周期的每个像素。该学生将测试假设，即SMFI可以通过数据驱动来识别环境压力的早期影响，从而使其能够应用于农业问题。最初的范例系统将以冬小麦品种和与营养缺乏相关的非生物胁迫为基础。第一个目标是提供人工智能（AI）或机器学习（ML）方法，将SMFI成像与植物早期营养缺乏定量联系起来。第二个目标是将这些信息与潜在的光合作用以及植物内部的初级和次级代谢联系起来。这个项目有实践和理论两个方面。该学生将使用正在进行的研究项目中开发的原型SMFI传感器和设备，设计和执行一系列受控植物胁迫试验。该学生将与生物学家一起工作，获得植物生理学的实用技能。数据收集将迭代地进行建模。该学生将研究并提供新的基于模型的人工智能方法，该方法可以适应和反卷积相互作用因素，并以闭环方式操纵光化（生长）光的调制，以最大限度地提高数据产生的信息内容。AI/ML方法正在成为解决许多物理建模或数据驱动逆问题的有吸引力的替代方案。深度神经网络（Deep neural networks, dnn）通过在多个尺度或层次上排列大量简单函数来构建复杂关系，并进一步结合抑制或子采样来实现灵活性和稳定性。生成对抗网络（GANs）是一种学习泛函的自适应框架。将生物模型注入深度学习将使深度学习更加集中和稳定。该项目将探索将植物生长中的生物机制（如光合作用、矿物质、营养物质和水分）整合到dnn中，以模拟其复杂的生长模式和功能的最佳方法。时间模型或机制也将通过约束或添加时间特征集成到建模网络中。将进行数据收集、模型开发和验证的对照试验。基于实验室研究的模型的预测和诊断能力将在温室和现场条件下进行测试。这项研究将填补信息科学、应用生物学和传感器系统研究的空白。通过利用所有三个学科的技术和方法知识，预计学生将开发一种“共同语言”，使他们能够与这些学科的博士和早期职业研究人员进行互动并影响他们。

Maintaining food security, with growing populations and changing climates, is one of the most important challenges facing humanity. In agriculture, this requires us to increase crop yields whilst reducing chemical inputs. Precision agriculture is one of the most promising and exciting approaches to address this, allowing targeted application of chemicals to minimise environmental impacts. This PhD project applies cutting edge AI approaches to understand plant nutrition and stress, using a novel sensor technology developed in Manchester - Sinusoidally Modulated Fluorescence Imaging (SMFI). SMFI generates big datasets (T-bytes) for every plant within a study comprised of a series of complex (phasors) harmonic parameters for every pixel at every fluorescence excitation period within an imaged crop sample. The student will test the hypothesis that SMFI can be data-driven to identify early impacts of environmental stress, allowing it to be applied to problems in agriculture. The initial exemplar system will be based on winter wheat species and abiotic stresses associated with nutrient deficiencies. The first aim is to deliver Artificial Intelligence (AI) or Machine Learning (ML) approaches that quantitatively relate SMFI imaging to early-stage nutrient deficiency in plants. A second aim is to link that information to the underlying photosynthesis, and primary and secondary metabolism within plants. This project has both practical and theoretical aspects. The student will design and execute a series of controlled plant stress trials, using prototype SMFI sensor and equipment developed in on-going research programmes. The student will work alongside biologists, gaining practical skills in plant physiology. Data collection will be performed iteratively for modelling. The student will research and deliver new model-based AI methodologies that can accommodate and deconvolute interacting factors as well as to manipulate the modulation of actinic (growth) light in a closed-loop manner, to maximise the information content of the data-arising. AI/ML methods are becoming an appealing alternative to solving many physical modelling or data driven inverse problems. Deep neural networks (DNNs) constructure a complex relationship using a large number of simple functions arranged in multiple scales or levels, further coupled with supressing or subsampling to achieve flexibility and stability. Generative adversarial networks (GANs) are an adaptive framework for learning universal functionals. Injecting biological models into deep learning will make deep learning more focused and stable. This project will explore the best ways to integrate biological mechanisms in plant growth such as photosynthesis, minerals, nutrients and water, into DNNs for modelling their complex growth patterns and functions. Temporal models or mechanisms will also be integrated to the modelling networks via constraints or adding temporal features. Controlled trials will be conducted for data collection, model development and verification. The predictive and diagnostic power of modelling based on laboratory studies will be tested under greenhouse and then field conditions.This research will bridge the gaps of information science, applied biology and sensor systems research. Through utilising knowledge of techniques and approaches from across all three disciplines it is anticipated that the student will develop a 'common language' allowing them to interact with and influence PhD and Early Career Researchers across the UoM community in those disciplines.