Abstract A range of modelling techniques were used in the estimation of dry matter content of intact mango fruit from short wave near infrared spectra, collected using an interactance geometry, with models developed on a data set collected across three seasons (n = 10,243) and tested on that of a fourth season (n = 1,448). Model types included Artificial Neural Network (ANN), Gaussian Process Regression (GPR), Local Optimized by Variance Regression (LOVR), Local Partial Least Squares Regression (LPLS), Local PLS Scores (LPLS-S) and Memory Based Learner (MBL), with manual tuning of parameters undertaken. Additionally, two commercially available cloud-based chemometric packages for automated model development were trialled. All of these models gave a better result than use of a global PLS model. The best result (lowest RMSEP) was achieved with an ensemble of ANN, GPR and LPLS-S, with the best individual model result achieved by LOVR, with RMSEP of 0.839 % and 0.881 %, respectively, compared to the global PLS result of 1.014 %. The best precision was achieved with the LPLS model, with a SEP of 0.846 %, compared to the global PLS result of 1.012 %. LOVR was twice as fast as a generalized latent variable selection method LPLS-S-cv in prediction of independent validation set (at 58.7 × 10−3 s compared to 163 x 10-3 s). The ANN model was satisfactory in all categories (prediction speed, model build speed, and prediction statistics) and insensitive to tuning, e.g., 33 of the 70 parameter combinations were within 0.05 units of RMSEP of the minimum combination. However, the ANN learning rate was low. For applications that require ‘real-time’ prediction, such as fruit packlines, use of ANN and GPR models is recommended. For non-cloud based handheld NIR devices lacking the computational power to perform local modelling, ANN is recommended, and LOVR or a model ensemble recommended in cloud based implementation. The automated cloud-based systems performed well (RMSEP of 0.850 % and 0.963 % for Hone Create and DataRobot, ensemble models, respectively), without human intervention for the choosing and tuning of models.

中文翻译：

---

为完整芒果果实干物质含量的 NIRS 模型实现跨季节、地点和栽培品种的稳健性。二、局部PLS和非线性模型

摘要 一系列建模技术用于从短波近红外光谱估计完整芒果果实的干物质含量，使用相互作用几何收集，在三个季节 (n = 10,243) 收集的数据集上开发模型并进行测试第四季 (n = 1,448)。模型类型包括人工神经网络 (ANN)、高斯过程回归 (GPR)、方差回归局部优化 (LOVR)、局部偏最小二乘回归 (LPLS)、局部 PLS 分数 (LPLS-S) 和基于记忆的学习器 (MBL) ，手动调整参数。此外，还试用了两种用于自动模型开发的商用基于云的化学计量学软件包。所有这些模型都比使用全局 PLS 模型给出了更好的结果。最好的结果（最低的 RMSEP）是用 ANN、GPR 和 LPLS-S 的集合实现的，LOVR 实现了最好的单个模型结果，RMSEP 分别为 0.839 % 和 0.881 %，而全局 PLS 结果为 1.014 %。LPLS 模型实现了最佳精度，SEP 为 0.846%，而全局 PLS 结果为 1.012%。在独立验证集的预测中，LOVR 是广义潜在变量选择方法 LPLS-S-cv 的两倍（与 163 x 10-3 s 相比为 58.7 × 10-3 s）。ANN 模型在所有类别（预测速度、模型构建速度和预测统计）中都令人满意，并且对调整不敏感，例如，70 个参数组合中的 33 个在最小组合的 RMSEP 的 0.05 个单位内。然而，ANN 学习率很低。对于需要“实时”预测的应用，例如水果包装线，建议使用 ANN 和 GPR 模型。对于缺乏执行本地建模的计算能力的非基于云的手持式 NIR 设备，建议使用 ANN，并在基于云的实现中建议使用 LOVR 或模型集成。基于云的自动化系统表现良好（Hone Create 和 DataRobot 集成模型的 RMSEP 分别为 0.850 % 和 0.963 %），无需人工干预来选择和调整模型。