Otsu method and its improved methods are widely suitable for the images whose histograms belong to Gauss distribution. However, they perform poor when the histograms of images belong to a mixture of other distributions. An algorithm based on the minimum class mean absolute deviation (MCMAD) is proposed. The new algorithm transforms two-dimension histogram into one-dimension histogram to decrease the computation complexity by diagonal projection method. The new proposed algorithm calculates the class mean and the class probability of every threshold in one-dimension histogram. The new proposed algorithm gets the minimum class mean absolute deviation of different thresholds by traversing all the thresholds in the one-dimension histogram. Among these thresholds, the threshold corresponding to the minimum class mean absolute deviation is the best segmentation threshold. Experimental results show that the new proposed algorithm not only can segment well on remote sensing images with histograms belonging to normal distribution, but also improve the performance of the remote sensing images with histograms belonging to Laplace distribution comparing with traditional Otsu method and its improved methods. Furthermore, the time consuming of the new proposed algorithm is low.