It brings about both higher computational complexity and data scale augmentation, while multi-directional information of remote sensing images has the superiority in extending function and increasing accuracy in cloud detection. Directonal information on cloud detection comes from BRDF of surface combined with atmospheric effects. Although determing certain view angles in local area by analytic solution is possible, but considering the complexities of application the traverse calculation for whole angles is still carried out in POLDER officer products. Due to redundant information existed on close neighbours of angler layer, averaged joint information entropy and K-L information divergence formed a feature basis for selection of angle subset. Two algorithms of optimal and idel solutions on Pareto multi-objectives front are proposed. Experiments of cloud detection were taken on two POLDER datasets firstly, then on a dataset of directional polarimetric camera （DPC） on board GF-5 satellite. The experimental results demonstrated that the overall accuracy of cloud detection by proposed method based on 2 angle- layers combinations is 89.36%, Kappa equals 0.7845 The validation on DPC dataset also showed that in comparison with GF-5 remote sensing synthetic image the similarity among them is 86%. The computation efficiency was raised to 7 times . Thus the proposed method takes advantages on computational cost saving and retaining multi-angle image’s intrinsic information. It contributes a new insight to cloud detection with its advantages of effectiveness, satisfactory accuracy and automatic operation.