College of Horticulture and Forestry Science, Tarim University, Alar, 843300, Xinjiang, People's Republic of China.
Institute of Mechanical Equipment, Xinjiang Academy of Agricultural Sciences, Shihezi, 832000, Xinjiang, People's Republic of China.
Sci Rep. 2024 Sep 6;14(1):20884. doi: 10.1038/s41598-024-71590-1.
The nitrogen content of apple leaves and jujube leaves is an important index to judge the growth and development of apple trees and jujube trees to a certain extent. The prediction performance of the two samples was compared between different models for leaf nitrogen content, respectively. The near-infrared absorption spectra of 287 apple leaf samples and 192 jujube leaf samples were collected. After eliminating the outliers by Mahalanobis distance method, the remaining spectral data were processed by six different preprocessing methods. BP neural network (BP), random forest regression (RF), least partial squares (PLS), K-Nearest Neighbor (KNN), and support vector regression (SVR) were compared to establish prediction models of nitrogen content in apple leaves and jujube leaves. The results showed that the determination coefficient (R), root mean square error (RMSE) and residual prediction deviation (RPD) of the models established by different combined pretreatment methods were compared among the five methods. Compared with the performance of the other four models, the modeling method of SG + SD + CARS + RF was suitable for the prediction of nitrogen content in apple leaves, and its modeling set R was 0.85408, RMSE was 0.082188, and RPD was 2.5864. The validation set R is 0.75527, RMSE is 0.099028, RPD is 2.1956. The modeling method of FD + CARS + PLS was suitable for the prediction of nitrogen content in jujube leaves. The modeling set R was 0.7954, RMSE was 0.14558, and RPD was 2.4264; the validation set R is 0.81348, RMSE is 0.089217, and RPD is 2.4552.In the prediction modeling of apple leaf nitrogen content in the characteristic band, the model quality of RF was better than the other four prediction models. The model quality of PLS in predictive modeling of nitrogen content of jujube leaves in characteristic bands is superior to the other four predictive models, These results provide a reference for the use of near-infrared spectroscopy to determine whether apple trees and jujube trees are deficient in nutrients.
苹果树和枣树叶片的氮含量是判断其生长发育的重要指标。分别比较了不同模型对苹果叶和枣树叶氮含量的预测性能。采集了 287 个苹果叶片样本和 192 个枣树叶样本的近红外吸收光谱。采用马氏距离法剔除异常值后，对剩余光谱数据分别采用 6 种不同的预处理方法进行处理。比较了 BP 神经网络（BP）、随机森林回归（RF）、最小二乘（PLS）、K-近邻（KNN）和支持向量回归（SVR），建立了苹果叶和枣树叶氮含量的预测模型。结果表明，采用不同组合预处理方法建立的模型的决定系数（R）、均方根误差（RMSE）和剩余预测偏差（RPD）在 5 种方法之间进行比较。与其他四种模型的性能相比，SG+SD+CARS+RF 组合建模方法更适合苹果叶氮含量的预测，其建模集 R 为 0.85408，RMSE 为 0.082188，RPD 为 2.5864。验证集 R 为 0.75527，RMSE 为 0.099028，RPD 为 2.1956。FD+CARS+PLS 组合建模方法更适合枣树叶氮含量的预测，建模集 R 为 0.7954，RMSE 为 0.14558，RPD 为 2.4264；验证集 R 为 0.81348，RMSE 为 0.089217，RPD 为 2.4552。在苹果叶特征波段氮含量预测建模中，RF 的模型质量优于其他四个预测模型。在枣树叶特征波段氮含量预测建模中，PLS 的模型质量优于其他四个预测模型，为利用近红外光谱法判断苹果树和枣树是否缺乏养分提供了参考。
