The existence of clouds in the atmosphere degrades the accuracy of aerosol retrieval. The empirical threshold method is popular in could detection, however its strong subjectivity and difficulty in coping with the dynamic spatial-temporal changes of the environment or the difference among satellite-borne detectors result in a large classification error at the boundary of ‘cloud’ and ‘clear’. In addition, its automatic detection is also poor. To achieve an effective detection of cloud over the land surface in the atmosphere, we propose a threshold optimized method which combines the statistical classification with data fusion of polarized multichannel remote sensing images. As for this method, a dual-brightness threshold to distinguish ‘cloud’ from ‘clear’ for most pixels is first derived based on the semi-supervised Kmeans clustering and its statistical features. Then, the joint confidence factor of multichannel data is calculated by the D-S evidence theory for each pixel in the fuzzy area of threshold neighborhood, and thus the fine threshold is acquired. The two objects of ‘cloud’ and ‘clear’ are finally and accurately classified in the sequential decision process. To validate the effectiveness of the proposed method, we perform a cloud detection experiment based on the remote sensing load data of POLRED3, and compare the measured results with the results of POLRED3. The results show that the classification by the proposed method is well consistent with that by the POLDER method with a high conformity of 95%. The error pixels are mostly located at the boundary between cloud and clear, indicating that the proposed method exhibits a favorable sensitivity to the classification at the cloud edge.