The polarization remote sensing experience threshold cloud detection algorithm is strongly affected by subjective factors， and it is very easy to have the problem of inaccurate cloud detection over bright ground. In response to this problem， this paper proposes a machine learning cloud detection algorithm that combines active and passive remote sensing satellites. The algorithm is based on the multi-channel multi-angle polarization characteristics of the POLDER3 payload and the high-precision cloud vertical characteristics of the CALIOP payload. It uses POLDER3 payload and CALIOP. The load observation overlaps the regional data， and the BP neural network optimized by the Particle Swarm Optimization algorithm is built to train the cloud detection model. Based on the cloud detection training model， a cloud detection experiment was carried out using POLDER3 level-1 data. The experiment showed that the cloud detection result of this algorithm is 92.46% consistent with the MODIS cloud detection product， which is higher than the consistency between the official POLDER3 cloud detection product and the MODIS cloud detection product 83.13%. By comparing the experimental results of the algorithm in this paper with the optical characteristics of different pixels from the official POLDER3 cloud detection product， it is found that compared with the official POLDER3 algorithm， this algorithm is more sensitive to thin clouds over the bright surface and can perform cloud detection more effectively.