In-season nitrogen (N) management is a promising strategy to achieve high tuber yield/quality and N use efficiency in potato (Solanum tuberosum L.) production. The SPAD-502 chlorophyll meter (SPAD) provides relative readings on plant N status using leaf chlorophyll transmittance and has the potential to replace the traditionally used expensive petiole analysis by estimating petiole nitrate-N (PNN) concentration non-destructively. The objective of this study was to develop a robust machine learning (ML) model for PNN concentration prediction across various genetic, environmental, and management conditions. Plot-scale experiments were conducted on an irrigated loamy sand soil in central Minnesota using a number of varieties and N fertilizer sources, application methods, and rates between 2010 and 2022. In each plot, approximately 20 petiole samples were collected for laboratory analysis, and 20 SPAD readings were collected and averaged. Weather information was collected by a nearby weather station. Three ML models (i.e. Random Forest, Extreme Gradient Boosting, and Support Vector Regression) were trained using Bayesian optimization in a nested 5-fold cross-validation. A near-linear trend was found between PNN concentration and the selected important features. Random Forest and Extreme Gradient Boosting regression models demonstrated that PNN concentrations could be predicted with an R2 of 0.8 using 15 features in a new site-year. When simplified by only using SPAD readings, cultivar information, accumulated growing degree days, accumulated total moisture, and as-applied N rates, these two tree-based models maintained the R2 values and achieved a 75 % diagnostic accuracy, outperforming both simple regression (66 %) and multivariate linear regression (70 %) models. We found that potato N status could be diagnosed accurately through PNN concentration prediction using chlorophyll meter and multi-source data fusion. The results of this study can be used as a baseline for future research on in-season N status diagnosis of potatoes involving different proximal and remote sensing technologies and N stress indicators.

中文翻译：

---

使用叶绿素计和机器学习的多源数据融合进行无损马铃薯叶柄硝酸盐-氮预测


当季氮 （N） 管理是在马铃薯 （Solanum tuberosum L.） 生产中实现高块茎产量/质量和氮利用效率的一种有前途的策略。SPAD-502 叶绿素仪 （SPAD） 使用叶绿素透射率提供植物氮状况的相对读数，并有可能通过无损估计叶柄硝酸盐 N （PNN） 浓度来取代传统上使用的昂贵的叶柄分析。本研究的目的是开发一个强大的机器学习 （ML） 模型，用于在各种遗传、环境和管理条件下预测 PNN 浓度。2010 年至 2022 年期间，在明尼苏达州中部的灌溉肥沃沙土上使用多种品种和氮肥来源、施用方法和施用量进行了小区规模的实验。在每个小区中，收集了大约 20 个叶柄样本用于实验室分析，并收集了 20 个 SPAD 读数并取平均值。天气信息是由附近的气象站收集的。三个 ML 模型（即 Random Forest、Extreme Gradient Boosting 和 Support Vector Regression）在嵌套的 5 折交叉验证中使用贝叶斯优化进行训练。在 PNN 浓度与所选重要特征之间发现近线性趋势。随机森林和极端梯度提升回归模型表明，在新站点年使用 15 个特征，可以以 0.8 的 R2 预测 PNN 浓度。当仅使用 SPAD 读数、品种信息、累积生长度天数、累积总水分和施用氮量进行简化时，这两个基于树木的模型保持了 R2 值并实现了 75% 的诊断准确性，优于简单回归 （66%） 和多变量线性回归 （70%） 模型。 我们发现，通过使用叶绿素计和多源数据融合预测 PNN 浓度可以准确诊断马铃薯 N 状态。本研究结果可作为未来马铃薯季节氮素状态诊断研究的基线，涉及不同的近感和遥感技术和氮素胁迫指标。