中文摘要：

近红外光谱（near infrared spectroscopy，NIRS）技术是一种快速、无损的仪器分析方法，在农产品品质检测方面引起了广泛的关注，在近红外光谱信息和品质指标之间建立一个稳健的模型是近红外光谱分析中十分重要且有一定难度的过程，常见的多元校正方法有偏最小二乘回归（PLSR）、主成分回归（PCR）和逐步多元线性回归（SMLR）等，该研究中除了常用的线性方法外，还采用了一种结合非线性方法的组合算法[结合了SMLR和径向基神经网络（RBFN）]用于梨坚实度的近红外光谱检测。比较常用的线性建模方法，原始光谱的PLSR模型的得到了较好的结果：校正集相关系数r=0．87，校正均方根误差RMSEC=3．88N，预测集r=0．84，预测均方根误差RMSEP=4．26N；组合算法的建模结果比SMLR和PCR的结果好，但比PLSR的结果稍差：校正集r=0．85，RMSEC=4．15N，预测集r=0．82，RMSEP=4．67N。结果表明：NIRS可用于梨的坚实度检测，但是建模方法的选择值得进一步研究以提高预测的精度。

英文摘要：

Near infrared （NIR） spectroscopy is an instrumental method, which was widely studied and used for rapid and nondestructive detection of internal qualities of agricultural products. Statistical modeling is a very important and difficult process in NIR detection to establish the relationship between nondestructive NIR spectral data and interested quality index of the products. Classical multivariate calibration methods such as partial least square regression （PLSR）, principle component regression （PCR）, stepwise multilinear regression （SMLR） were often used for modeling. In the present study, besides these algorithms, another mixed algorithm was adopted for establishing a nonlinear model of NIR spectra and Magness Taylor（MT） firmness of ＂Xueqing＂ pears. The mixed algorithm was combined with SMLR and artificial neural network （ANN）. NIR diffuse reflectance spectra of intact pears were measured in the spectral range of 800-2 630 nm using InGaAs detector. However, only spectral information between 800 and 2 500 nm was used for modeling because of the low signal to noise ratio beyond 2 500 nm. Comparing the classical multivariate calibration methods of PLSR, PCR and SMLR, the modeling results using PLSR method were much better than the other two methods. Moreover, models based on original spectra turned out better results than models based on derivative spectra for all the three methods. The best results were r=0. 87, RMSEC=3.88 N of calibration and r=0. 84, and RMSEP=4. 26 N of validation by using PLSR method based on original spectra. The mixed algorithm also performed better than SMLR and PCR, but was a bit worse than PLSR： r=0. 85, RMSEC=4. 15 N of calibration and r=0. 82, and RMSEP=4. 67 N of validation. The results indicated that fruit NIR spectra could be used for MT-firmness prediction when a proper algorithm was chosen, however, further study on statistic modeling is still necessary to improve the predicting performance.