Abstract: Phosphorus is very important for the growth and development of rubber trees. There are many studies on the rapid non-destructive detection of nitrogen, phosphorus and potassium content of rubber trees based on near infrared spectroscopy, but the detection effect of phosphorus content is far less than that of nitrogen and potassium. Based on near infrared hyperspectral technology, combined with nonlinear feature extraction method and modeling algorithm, the rapid non-destructive detection of phosphorus content in rubber tree was realized. The near-infrared hyperspectral data of rubber leaves were taken as the analysis object, and the feature extraction idea of neural network nonlinear transformation was used to construct a model fusing autoencoder and feedforward neural network (AE-FFNN). The nonlinear spectral feature information of rubber leaves was extracted by autoencoder, and the model was established by using feedforward neural network to deal with classification tasks with different levels of fineness to realize the qualitative analysis of phosphorus content in rubber leaves. The results showed that AE-FFNN model could effectively extract spectral nonlinear features and compress the number of features. 31 features were extracted by this method, and the accuracy of qualitative analysis model was improved, which reached to 91.10%. Compared with the machine learning model widely used in the field of spectral detection, the performance of the AE-FFNN model was greatly improved. The model could be used not only for qualitative analysis of phosphorus content in rubber leaves, but also for quantitative study of phosphorus content.

Key words: phosphorus, rubber tree leaves, autoencoder, feedforward neural network, classification

摘要： 磷元素对橡胶树的生长发育至关重要，基于近红外光谱的橡胶树氮磷钾元素含量的快速无损检测已有很多研究，但磷元素含量的检测效果远不如氮钾。基于近红外高光谱技术结合非线性特征提取方法和建模算法实现橡胶树磷元素含量的快速无损检测。以橡胶树叶片的近红外高光谱数据为分析对象，运用神经网络非线性变换的特征提取思想，构建了一种融合自编码器与前馈神经网络(autoencoder-feedforward neural network, AE-FFNN)模型。通过自编码器提取橡胶树叶片的光谱非线性特征信息，运用前馈神经网络进行建模，应对不同精细程度的分类任务，从而实现橡胶树叶片磷含量的定性分析。结果表明，AE-FFNN模型有效提取了光谱非线性特征并压缩了特征数量，通过该方法提取的特征为31个，且定性分析模型精确度提升，能够达到91.10%。相较于在光谱检测领域广泛采用的机器学习模型，所建立的AE-FFNN模型性能有较大提升。该模型既可以应用于橡胶树叶片磷元素含量的定性分析，也可为磷元素含量的定量研究提供思路。

关键词: 磷含量, 橡胶树叶片, 自编码器, 前馈神经网络, 分类