Mango DMC and NIR spectra

The dry matter content of mango fruit is an important metric for determining harvest maturity and ensuring the eating quality of the ripened fruit. Near infrared spectroscopy can be used as a non-invasive method of estimating attributes of individual fruit, including dry matter content. The technique relies on statistic models (‘chemometrics’) to deduce information on sample attributes from spectra collected from the fruit. Barriers to the adoption of this technique for practical use in the fruit industry include the robustness of models across fruit from different growing conditions and spectra collected on different instruments. The proposed research is intended to reduce these barriers for the assessment of mango dry matter content by exploring new techniques for developing robust, global models across season, growing conditions, fruit variety, individual instruments and other variations. This would allow new instruments to be used ‘out of the box’ without the need for local calibration, hence greatly reducing the cost of uptake. Deep learning modelling techniques have been recently applied to spectroscopic applications, with claims of improved performance over the standard chemometric method, Partial Least Squares Regression, although these studies have typically involved relatively small datasets with limited testing on new populations of data. With access to an extended dataset of over 80,000 spectra from over 500 fruit populations, the aim of this proposed study is to validate previous publication claims that the use of a Convolutional Neural Network (CNN) model, a deep learning technique, is superior to existing methods in NIRS based prediction of mango dry matter content. The study also aims to optimise the operation and the architecture of the CNN model over that employed in previous publications, in context of the mango dry matter data set and use on a portable instrument.

芒果果实的干物质含量是判断采收成熟度、保障成熟果实食用品质的重要指标。近红外光谱（Near Infrared Spectroscopy）可作为非侵入式检测手段，用于估测单果的多项属性，其中便包括干物质含量。该技术依托统计模型（化学计量学，chemometrics），从采集得到的果实光谱数据中推导样本属性的相关信息。该技术在果品产业实际应用中面临的核心障碍，在于不同种植条件下果实样本的模型鲁棒性，以及不同仪器采集的光谱间的适配性不足。本拟开展的研究旨在攻克上述障碍，针对芒果干物质含量的检测需求，探索可在不同季节、种植环境、果实品种、仪器设备及其他变量间构建稳健通用模型的新方法。这一目标一旦实现，新仪器便可“开箱即用”，无需进行本地校准，从而大幅降低技术推广的应用成本。近年来，深度学习建模技术已被应用于光谱分析领域，相关研究声称其性能优于经典化学计量学方法——偏最小二乘回归（Partial Least Squares Regression），但此类研究通常采用的数据集规模相对有限，且在新数据群体中的验证较为不足。依托涵盖500余个果实群体、超过80000条光谱的大型扩展数据集，本拟开展的研究旨在验证此前已发表的研究结论：相较于现有方法，卷积神经网络（Convolutional Neural Network, CNN）这一深度学习技术，在基于近红外光谱的芒果干物质含量预测任务中表现更优。本研究同时还将针对芒果干物质数据集场景与便携式仪器的使用需求，对此前研究中采用的CNN模型的操作流程与架构进行优化。