基于光谱特征分析的苹果树叶片营养素预测模型构建

该文旨在利用光谱分析技术建立高精度苹果叶片营养素预测模型，为苹果树的精细管理提供技术支持。在苹果树年度生长周期的坐果期、生理落果期和果实成熟期等重要物候期，采集了180个果树叶片样本并测量了果树叶片在可见光和近红外波段的反射光谱，同时在实验室采用化学方法获取了果树叶片的氮素以及叶绿素含量。对于聚类后样本，分别分析了果树叶片反射光谱以及经小波滤波后的反射光谱与叶绿素以及氮素之间的相关关系，而后利用偏最小二乘和支持向量机（SVM，support vector machine）方法分别建立了果树叶片叶绿素和氮素含量的回归模型。研究发现，随着生长阶段的推进，在可见光处的反射率逐渐升高，在近红外处的反射率逐渐降低，且基于小波滤波反射光谱的营养素SVM回归模型精度最高：建立的叶绿素回归模型，其测定系数R2达到0.9920，均方根误差 RMSE为0.0039，验证精度R2达到0.9036，RMSE为0.1979；建立的氮素回归模型，其测定R2和验证R2也达到0.74以上，模型的回归RMSE为0.0554，验证RMSE为0.1215。结果表明，采用支持向量机回归模型可以精确估计果树叶片叶绿素含量，对氮素含量的估计精度也达到了实用化水平。

Construction of apple tree leaves nutrients prediction model based on spectral analysis

Abstract: This research aimed at exploring the VIS/NIR (Visible Spectrum/ Near Infra Red) reflectance spectral characteristics of apple tree leaves, and establishing a high-precision model to predict nutrient content for these leaves. Samples were collected from the apple orchard of Beijing Xiangtang culture village during the period of fruit-bearing, fruit-falling and fruit-maturing separately. The apple trees in the orchard were in the full productive age. Twenty apple trees (15 year-on trees and 5 year-off trees) were selected randomly from different regions. Then a main branch of each target tree was selected and three representative parts (base part, middle part and top part) of every bough were marked, and the leaves from the same representative part were considered as one sample. In the end, 60 samples of apple leaves were collected in each phenological period, and the visible and near infrared spectral reflectance were measured using Shimadzu UV-2450 spectrograph. At the same time, the chlorophyll content for each sample was detected using spectrophotometry and the nitrogen content of each sample was measured using the Kjeldahl method in the laboratory. The obtained spectral reflectance and nutrient content were clustered based on each bough individually. The first through seventh layer wavelet decompositions were done to the original spectrum respectively. It can be perceived that with the decomposition scale increasing, the curve became smoother because of eliminating the impact of random noise, while some valid information was lost at the same time. According to the correlation analysis, this study selected 3-layer db4 wavelet filtering spectral information to predict the nitrogen and chlorophyll content. After that, correlation analyses were conducted between: 1) the chlorophyll content of apple tree leaves and their spectral reflectance; 2) the chlorophyll content of apple tree leaves and their spectral reflectance under wavelet filtering; 3) the nitrogen content of apple tree leaves and their spectral reflectance; and 4) the nitrogen content of apple tree leaves and their spectral reflectance under wavelet filtering. Then, the regression models for predicting nitrogen content and chlorophyll content of apple tree leaves were established using PLS (Partial Least Square) and SVM (Support Vector Machine) methods, respectively, based on the above spectral signal. The results indicated that: 1) with the advance of growth stage, the reflectance at visible waveband increased gradually, while at the near infrared waveband, the reflectance decreased gradually; 2) wavelet analyzing technology could distinguish the mutation part and noise in the spectral signals effectively, which make it possible to retain the maximum amount of effective information during the signal denoising process. The wavelet filtering technology played a significant role in promoting the modeling accuracy in predicting the Chlorophyll; 3) the models based on the SVM method had higher accuracies; 4) for the Chlorophyll regression model based on the spectral reflectance under wavelet filtering, the calibration R2 reached to 0.9841, RMSEC was 0.0039, and the validation R2 of reached to 0.9036, RMSEP was 0.0567; and 5) for the nitrogen regression model, the R2 of calibration and validation model were all above 0.74, RMSEC was 0.0554 and RMSEP was 0.1215. It was concluded that the chlorophyll SVM regression model reached a high accuracy, and the nitrogen SVM regression model also reached the practical level with high stability.