基于近红外特征光谱的番茄苗氮含量快速测定方法研究

为了提高近红外光谱技术快速测定番茄苗氮含量的准确度和稳健性，比较分析竞争自适应重加权采样法(CARS)、蒙特卡罗无信息变量消除法(MCUVE)、向后间隔偏最小二乘法(BiPLS )和组合间隔偏最小二乘法(SiPLS)四种特征波长挑选方法，筛选与番茄苗氮含量相关的特征光谱。在十种不同氮素处理水平下(尿素溶液浓度0～120 mg·L-1)，培育60株番茄苗样本(每个处理6株)，使其分别处于不同程度的过量氮素、氮素适度、缺氮素和无氮素状态。分别采集每株番茄苗样本的叶片，扫描其12 500～3 600 cm-1波段的近红外光谱。比较四种方法所建立的番茄苗氮素定量分析模型可知：CARS和MCUVE挑选的特征变量所建定标模型的性能比BiPLS和SiPLS挑选的特征变量所建定标模型的性能更优，但是预测性能远低于后者。其中，基于BiPLS建立的番茄苗氮素含量预测模型性能最佳，相关系数(r)、预测均方根误差(RMSEP)和性能对标准差之比(RDP)分别为0.952 7，0.118 3和3.291 0。因此，近红外光谱技术结合特征谱区筛选可以有效地提高番茄苗叶片氮素含量的定量分析模型指标，使模型更实用化。但是，特征波长挑选方法不具有普适性。基于单个波长变量筛选的方法所建立的模型较为敏感，更适用于样本状态较为均匀的待测对象;而基于波长区间筛选的方法所建的模型相对抗干扰性更强，更适用于样品状态不均匀，重现性较差的待测对象。因此，特征光谱筛选只有与样本状态及建模指标结合，才能使其在建模过程中发挥更好的作用。

Research on Fast Detecting Tomato Seedlings Nitrogen Content Based on NIR Characteristic Spectrum Selection

In order to improve the accuracy and robustness of detecting tomato seedlings nitrogen content based on near-infrared spectroscopy (NIR), 4 kinds of characteristic spectrum selecting methods were studied in the present paper, i.e. competitive adaptive reweighted sampling (CARS), Monte Carlo uninformative variables elimination (MCUVE), backward interval partial least squares (BiPLS) and synergy interval partial least squares (SiPLS). There were totally 60 tomato seedlings cultivated at 10 different nitrogen-treatment levels (urea concentration from 0 to 120 mg·L-1)，with 6 samples at each nitrogen-treatment level. They are in different degrees of over nitrogen, moderate nitrogen, lack of nitrogen and no nitrogen status. Each sample leaves were collected to scan near-infrared spectroscopy from 12 500 to 3 600 cm-1. The quantitative models based on the above 4 methods were established. According to the experimental result, the calibration model based on CARS and MCUVE selecting methods show better performance than those based on BiPLS and SiPLS selecting methods, but their prediction ability is much lower than that of the latter. Among them, the model built by BiPLS has the best prediction performance. The correlation coefficient (r), root mean square error of prediction (RMSEP) and ratio of performance to standard derivate (RPD) is 0.952 7, 0.118 3 and 3.291, respectively. Therefore, NIR technology combined with characteristic spectrum selecting methods can improve the model performance. But the characteristic spectrum selecting methods are not universal. For the built model based on single wavelength variables selection is more sensitive, it is more suitable for the uniform object. While the anti-interference ability of the model built based on wavelength interval selection is much stronger, it is more suitable for the uneven and poor reproducibility object. Therefore, the characteristic spectrum selection will only play a better role in building model, combined with the consideration of sample state and the model indexes.