According to the information redundancy caused by high classification feature space dimension, we have proposed a method of cloud detection based on linear dimension compression for feature space. First, the classification feature parameters were extracted and the feature space was established. According to the different distribution structures of cloud and underlying surface samples, based on the completeness of classified information in compressed subspace, the sample’s probability distribution model is established. Then the model parameters are solved by using the maximum likelihood estimation method and the optimum transformation matrix is estimated, which is used for features compression and decorrelation. At last, a classifier is introduced for cloud detection in the compressed subspace. Experimental results have shown that the proposed method can effectively remove the classification features redundancy of the clouds and underlying surfaces, and achieve effective separation of clouds and underlying surfaces in compressed subspace. The cloud detection probability in optical remote sensing images is up to 98％ or more.