Hyperspectral (HS) images due to the simultaneous data acquisition in hundreds of narrow and close spectral bands, have high between bands correlation. In order to storage and transformation, they need to be compressed. A number of lossy/lossless methods have been developed for data compression in spatial or spectral domain. Spectral information of HS data has much more of importance than spatial information; therefore compression should be done in such a way that the spectral information is well preserved. In this paper, a lossy compression technique in the spectral domain is proposed by using curve fitting. The method has better performance compared to data compressing method using principal component analysis. In the presented method, the spectral reflectance curve (SRC) of each pixel is divided into a few non-overlapping intervals based on a specific criterion, and then, a polynomial function is fitted on each interval. The calculated coefficients of each fitted curve are considered as the new features of that section of the SRC. The experimental results show that the compressed data after the recovery is very similar to the original data.