A Compressive Sensing(CS) multiple description coding scheme with hybrid sampling was proposed to improve the coding efficiency of the traditional CS coding system and to maintain the ability of resisting packet loss. In the scheme, both 2-D Discrete Cosine Transformation( DCT) matrix and sub-Gaussian matrix were used to measure the image signal simultaneously. Then, a Golomb code and its improved version were used to encode for the resulted measurements, respectively. As a result, the 2-D DCT measurement bit streams with complete code words and the Gaussian measurement bit streams with incomplete code words were obtained respectively. In the decoder, these incomplete code words could be decoded successfully with a Maximum A posteriori Probability (MAP) estimator, and the deficient code words could be estimated by the relevance between 2-D DCT and Gaussian measurements. Finally, these decoded measurements were grouped together again to reconstruct the image signal by solving a 1-norm optimization problem. Experimental results on both natural and remote sensing images show that the Peak Signal to Noise Ratio(PSNRs) of the images reconstructed by proposed method can be superior to that of traditional CS coding scheme by 2~4 dB at different packet loss rates, meanwhile, it has a robust resisting packet loss ability.