总磷（TP）、总氮（TN）是水质富营养化的重要指标，亦是水质监测的主要参数。光谱法水质监测因其快速高效、无二次污染等特点，成为当前遥感水环境研究的热点。通常的TP、TN反演模型或是基于实验室配置标准液进行光谱测量而建立的，或是基于采集的全样本进行建模，这种方式构建的模型虽具有较好的回归效果，但水体中由于各种水质参数的相互影响，TP、TN的浓度分布不均匀，预测值存在高于或低于建模样本浓度的可能，使得实际的预测效果并不理想。鉴于此，本文以白洋淀实验区的实际水体样本为反演模型的输入值，在确定最优相关波段和最佳反演模型的基础上，讨论了不同浓度范围的样本建模对反演模型的影响，确定了模型对超出建模浓度值样本的预测能力。通过5种不同方式的建模和预测分析，其结果表明：建模样本浓度覆盖预测样本时，反演模型决定系数R2＞0.6，TP、TN浓度预测值的平均偏离度ARE＜20%；建模浓度高于预测样本时，R2在0.6左右，对超过建模浓度范围12%以内的预测值，其ARE＜25%；建模浓度低于预测值时，R2介于0.4~0.5，预测值超过建模样本浓度一倍时，ARE≤30%；建模样本浓度位于预测值两侧时，R2可达0.8，ARE＜25%；建模样本浓度值介于预测值之间时， 0.45＜R2＜0.55，ARE＞35%。通过本文的研究与讨论，可为水质参数监测的实际工程应用提供科学依据和参考。

Objective, Total phosphorus (TP) and total nitrogen (TN) are important indicators of eutrophication of water quality and are also the main parameters for water quality monitoring. Water quality monitoring by spectroscopy has become a hot spot in the current remote sensing water environment research due to its rapid, efficient, and no secondary pollution. The usual TP and TN inversion models are established based on the laboratory configuration standard solution for spectral measurement, or modeling based on the full samples. Although the model constructed in this way has a good regression effect,but the actual water body due to the mutual influence of various water quality parameters, the concentration distribution of TP and TN is not uniform, and the predicted value may be higher or lower than the concentration of the modeling sample, making the actual prediction effect unsatisfactory.Method, In this paper, the actual water samples in the Baiyangdian area are used as the input values of the inversion model. First, the measured spectral data and the chemical analysis values of TP and TN are used to compare the relationship between the calculated values of different reflectances and water quality parameters; A variety of inversion models have been constructed for the best relevant bands , and the most stable and accurate modeling method has been determined through comparison; on this basis, the modeling samples are divided into uniform samples, high value samples, low value samples, median value sample and two side value samples according to their concentration .Then discussed the influence of sample modeling with different concentration ranges on the inversion model, and determined the model""s predictive ability for samples with concentration values beyond modeling.Result, From the extraction results of the characteristic wavebands, in the range of 400-100nm, the reflectance correlation coefficient of TP and TN corresponding to single wavelength is less than 0.3, which is not high; the maximum correlation coefficient with the first-order value of reflectance is 0.76, which is a moderate correlation; the correlation coefficients with the reflectance ratio are all over 0.8, which is highly correlated. In the inversion effect of the linear regression model, the exponential method and the logarithmic method are inferior to the multiple power method, and the effect of high power is better than low power, but the overall effect is not ideal, the all model R2 is less than 0.6. Concentration of modeling sample when the range is different, the R2 of the model is also different. The result is: two-sided method> uniform method> high value method> middle method> low value method. When the modeled sample concentration covers the predicted sample, the inversion model determination coefficient R2>0.6, the average deviation of the predicted value(ARE) of TP and TN concentration ARE<20%; when the modeled concentration is higher than the predicted sample, R2 is about 0.6. The predicted value within 12% of the over concentration range has an ARE<25%; when the modeling concentration is lower than the predicted value, R2 is between 0.4 and 0.5, and when the predicted value exceeds the modeling sample concentration, the ARE≤30%; When the sample concentration is on both sides of the predicted value, R2 can reach 0.8 and ARE<25%; when the modeled sample concentration is between the predicted value, 0.4535%.Conclusion, When using the reflectance method for TP and TN inversion, the ratio method can be given priority to the characteristic band when modeling; the regression effect of the partial least square(PLS) method is significantly better than that of multiple power and exponential models, and the model also has a clear physical meaning, so it can be used in the regression study of TP and TN based on reflectance; for the predicted value that is not within the range of the modeled sample concentration, the credibility of the inversion results can be judged based on the relative relationship between its concentration value and the modeled sample concentration value.