The development of compressive techniques of hyperspectral images (HSI) becomes critical due to the limitations of the requirements of data storage, transmission and processing. Recently, Compressed Sampling (CS) has been used for hyperspectral imaging for its ability to recover the original data exactly under certain condition at a much lower sampling rate than Nyquist rate. In this paper, a residual reconstruction algorithm incorporating with two dimensional compressed sampling (2DCS) for Hyperspectral images is proposed to improve the performance of reconstruct algorithm. In the reconstruction process, spectral prediction is introduced for the strong spectral correlation between hyperspectral image bands. The experimental results reveal that the proposed technique achieves significantly higher quality than a straightforward reconstruction that reconstructs the hyperspectral images band by band independently. Meanwhile, the comparison between 2DCS and block-based compressed sampling (BCS) is developed and the results demonstrate that the superiority of 2DCS over BCS is in terms of high PSNR with respect to sampling rate.