This research was conducted to develop a method to remove the effect of surface temperature of Shingo pears for sugar content measurement. Sugar content was measured by a near-infrared spectrum analysis technique. Reflected spectrum and sugar content of a pear were used for developing regression models. For the model development, reflected spectrums having wavelengths in the range of 654 to 1,052nm were used. To remove the effect of surface temperature, special sample preparation techniques and partial least square (PLS) regression models were proposed and tested. 71 Shingo pears stored in a cold storage, which had 2$^{circ}C$ inside temperature, were taken out and left in a room temperature for a while. Temperature and reflected spectrum of each pear was measured. To increase the temperature distribution of samples, temperature and reflected spectrum of each pear was measured four times with one hour twenty minutes interval. During the experiment, temperature of pears increased up to 17 $^{circ}C$. The total number of measured spectrum was 284. Three groups of spectrum data were formed according to temperature distribution. First group had surface temperature of 14$^{circ}C$ and total number of 51. Second group consisted of the first and the fourth experiment data which contained the minimum and the maximum temperatures. Third group consisted of 155 data with normal temperature-distribution. The rest data set were used for model evaluation. Results shelved that PLS model I, which was developed by using the first data group, was inadequate for measuring sugar content of pears which had different surface temperatures from 14$^{circ}C$. After temperature compensation, sugar content predictions became close to the measured values. Since using many data which had wide range of surface temperatures, PLS model II and III were able to predict sugar content of pears without additional temperature compensation. PLS model IV, which included the surface temperatures as an independent variable. showed slightly improved performance(R$^2$=0.73). Performance of the model could be enhanced by using samples with more wide range of temperatures and sugar contents.