A machine learning approach to predicting agro-chemical transport in soil using X-ray imaging

在政府的产业战略：重大挑战政策文件(2017)中，机器学习被强调为人工智能的优先事项。深度机器学习(一种人工智能，允许计算机从大数据集中快速学习，而无需明确编程)，尤其是卷积神经网络(CNN)，近年来同样给计算机视觉带来了革命性的变化，提供了能够从复杂图像中进行高级模式识别的工具。CNN成功的关键是它们同时学习如何识别关键图像特征并说明它们在整个模式中的重要性的能力。CNN的一个关键特征是能够识别CNN学习使用的图像中的哪些形态特征来生成所需的网络。同样，X射线成像在土壤和农业科学中的应用也彻底改变了我们在3D中探索土壤结构的能力，就像它在野外存在的那样。如孔隙连通性等重要属性可从图像中得出，并可用于说明土壤的关键功能，如保水(即干旱)和水力传导性(即洪水风险)。目前取得进展的一个障碍是，收集的海量数据往往超过当地的计算能力，因此，通常只利用一小部分可用数据。与先正达(世界上最大的农用化学品公司之一)合作，我们将使用从X射线图像开发的CNN来显示土壤结构影响农用化学品在土壤中的运输。这将使我们能够告知管理实践，以确保土壤中保留更多的化肥和农药(目前有75%的损失是由于环境成本&gt；GB 1B/年)。更有针对性地在土壤中应用农用化学品对英国公司和公民的好处将包括更清洁的水，在环境修复上花费的更少的GB，以及改善土壤健康，从而提高作物产量。

Machine Learning is highlighted as a priority as part of Artificial Intelligence in Government's Industrial Strategy: Grand Challenges Policy Paper (2017). Deep machine learning (a type of AI that allows computers to learn rapidly from large datasets without being explicitly programmed), and convolutional neural networks (CNNs) in particular, have similarly revolutionised computer vision in recent years, providing tools capable of high level pattern recognition from complex images. The key to CNNs success is their ability to simultaneously learn both how to identify key image features and account for their importance within the pattern as a whole. A key feature of CNNs is the ability to identify what morphological features within an image the CNN learned to use to generate the required network. In a similar way the application of X-ray imaging to soil and agricultural sciences has revolutionised our ability to explore soil structure in 3D, as it exists in the field. Important properties such as pore connectivity can be derived from images and used to account for crucial soil functions such as water retention (i.e. drought) and hydraulic conductivity (i.e. flood risk). A current impediment to progress is the vast amount of data collected often exceeds local computing capacity, and as such, it is common that only a small amount of the available data is exploited. Partnering with Syngenta (one of the World's largest agro-chemical companies) we will use CNNs developed from X-ray images to show soil structure influences transport of agrochemicals through soils. This will enable us to inform management practices that could be used to ensure more fertiliser and pesticide is retained in soil (>75% is currently lost to the environment costing >£1B/year). The benefits to UK plc and citizens from a more targeted application of agro-chemicals to soil would include cleaner water, less £ spent on environmental remediation and improved soil health leading to higher crop yields.