中文摘要：

以温室栽培黄瓜作物为对象，分析了叶片反射光谱与叶片磷素含量之间的相关关系，并建立了预测模型。首先利用便携式光谱辐射仪测量了自然光照条件下温室黄瓜叶片的光谱反射率，并计算了反射率光谱的一次微分光谱。相关分析表明反射率光谱与叶片含P量之间具有一定的相关性，但线性相关不显著。利用微分光谱可以部分消除系统误差、背景噪声等的影响，明显提高了相关系数，但预测模型精度仍然达不到实用程度。在理论分析的基础上，选取978，920，737和458nm等4个波长作为特征波长，分别利用人工神经网络和支持向量机建立了黄瓜叶片P素含量对应于微分光谱特性的非线性预测模型，结果表明两种算法都获得了较好的预测效果，支持向量机模型的预测能力（Rv=0．754）优于人工神经网络模型（Rv=0．712）。

英文摘要：

A handheld spectroradiometer was used to measure the spectral reflectance of the crop with the measurable range from 325 nm to 1 075 nm. Since the first derivative of the spectra can well eliminate spectral error, it was caleulated for each spectrum. The cucumber leaves were also sampled and the phosphorus content was measured for each sample with chemical method. First, the correlation between the phosphorus content of the cucumber leaf and the spectral reflectance was analyzed but high coefficient was not obtained. It was shown that there is not high linear relation between those. Then, the analysis was conducted between the phosphorus content of the cucumber leaf and the first derivative of spectrum for each sample. The coefficients were improved. However, it was not high enough to establish an estimation model. It shows that non-linear model is needed to estimate the phosphorus content of the crop leaf based on spectral reflectance. Artificial neural network （ANN） and support vector machine （SVM）, the modern ealgorithm for modeling and estimating, were used to establish the nonlinear models. From stepwise multi-regression, four wavelengths, 978, 920, 737 and 458 nm, were selected as modeling wavebands. For the Artificial Neural Network （ANN） model, the data of spectral reflectance in the four wavebands were taken as the input and the phosphorus content was taken as the output. And the number of the neurons in the middle layer, the learning rate, and the learning error were set as 25, 0. 05, and 0. 001, respectively. The calibration accuracy of the model was 0. 995, and the validation accuracy reached to 0. 712. For the Support Vector Machine （SVM） model, the selected kernel function was anova, and the penalty parameter C and the linear e-insensitive loss function were set as 100 and 0. 000 01, respectively. The calibration accuracy of the model was closed to 1, and the validation accuracy reached to 0. 754. It can be concluded that both nonlinear models are praetical.