Hssaini Lahcen, Razouk Rachid
Agro-Food Technology and Quality Laboratory, Regional Center of Agricultural Research of Meknes, National Institute of Agricultural Research, Rabat, Morocco.
Anal Sci. 2025 Jun 29. doi: 10.1007/s44211-025-00814-9.
Predicting leaf mineral composition is critical for monitoring plant health and optimizing agricultural practices. This study combines Fourier-transform infrared spectroscopy with attenuated total reflectance (FTIR-ATR) and machine learning (ML) to specific macronutrients, namely nitrogen (N), phosphorus (P), potassium (K), calcium (Ca), and magnesium (Mg), in fig leaves (Ficus carica L.). A dataset of 90 leaves was analyzed, with FTIR spectra (450-4000 cm⁻) preprocessed via baseline correction and second-derivative transformations. Three ML models were evaluated using fivefold cross-validation including Random Forest (RF), Support Vector Regression (SVR), and Gradient Boosting (GB), with performance assessed via root mean square error (RMSE), coefficient of determination (R), and ratio of performance to deviation (RPD). GB outperformed other models, achieving validation RMSE/RPD values of 0.133/1.60 (nitrogen, N), 0.0107/1.79 (phosphorus, P), 0.1328/1.65 (potassium, K), 0.0636/1.96 (magnesium, Mg), and 0.2657/1.60 (calcium, Ca). Predictions for Mg (validation R = 0.7351) and P (validation R = 0.6873) exhibited the highest accuracy, potentially attributed to their stronger or more distinct spectral features (e.g., Mg-O stretching around 1050- 1150 cm⁻; P-O vibrations around 1240 cm⁻). Cross-validation revealed robust generalization for GB; while mean training RMSE was very low (< 0.01 for P and Mg), validation RMSE remained relatively low, underscoring the model's utility for screening (RPD > 1.5). Despite evidence of overfitting (training R ≈ 0.999 vs. validation R = 0.61-0.74), GB's performance evaluated using both RMSE and RPD confirmed its superiority over RF and SVR, which showed higher errors (e.g., SVR for Ca: RMSE = 0.4574, RPD = 1.07). This study demonstrates that FTIR-ATR coupled with ML is a rapid, non-destructive alternative to conventional destructive chemical analysis and that GB's reliability, as indicated by RPD values > 1.5, offers actionable insights for precision nutrient management in sustainable agriculture.
预测叶片矿物质组成对于监测植物健康状况和优化农业实践至关重要。本研究将傅里叶变换红外光谱与衰减全反射（FTIR-ATR）以及机器学习（ML）相结合，用于测定无花果叶（Ficus carica L.）中特定的大量营养素，即氮（N）、磷（P）、钾（K）、钙（Ca）和镁（Mg）。分析了一个包含90片叶子的数据集，其FTIR光谱（450 - 4000 cm⁻）通过基线校正和二阶导数变换进行预处理。使用五折交叉验证对三种ML模型进行了评估，包括随机森林（RF）、支持向量回归（SVR）和梯度提升（GB），通过均方根误差（RMSE）、决定系数（R）和性能与偏差比（RPD）来评估性能。GB的表现优于其他模型，在验证中氮（N）的RMSE/RPD值为0.133/1.60，磷（P）为0.0107/1.79，钾（K）为0.1328/1.65，镁（Mg）为0.0636/1.96，钙（Ca）为0.2657/1.60。镁（验证R = 0.7351）和磷（验证R = 0.6873）的预测准确率最高，这可能归因于它们更强或更明显的光谱特征（例如，1050 - 1150 cm⁻附近的Mg - O伸缩振动；1240 cm⁻附近的P - O振动）。交叉验证显示GB具有强大的泛化能力；虽然平均训练RMSE非常低（P和Mg < 0.01），但验证RMSE仍然相对较低，这突出了该模型在筛选方面的实用性（RPD > 1.5）。尽管有过拟合的迹象（训练R ≈ 0.999，而验证R = 0.61 - 0.74），但使用RMSE和RPD评估的GB性能证实了它优于RF和SVR，后两者显示出更高的误差（例如，钙的SVR：RMSE = 0.4574，RPD = 1.07）。本研究表明，FTIR-ATR与ML相结合是一种快速、无损的传统破坏性化学分析替代方法，并且GB的可靠性（RPD值> 1.5表明）为可持续农业中的精准养分管理提供了可操作的见解。
