Abstract: Abstract: This study Aiming at the automatic observation of ground-based cloud cover in complex meteorological scenarios, this paper proposes a dynamic multi-threshold cloud cover identification method driven by solar altitude and visibility. Based on a database of 190,628 ground-based cloud image samples from 24 representative stations across China from 2021 to 2023, a machine learning approach is employed to construct an adaptive multi-threshold generation model by fusing visible light radiation characteristics of ground-based cloud images, real-time meteorological optical range, and solar altitude angle parameters. This model breaks through the adaptability bottleneck of traditional fixed threshold methods in scenarios such as dawn/dusk, thin clouds, and low visibility. The results show that the cloud-sky segmentation threshold increases with the rise of solar altitude angle, with threshold ranges of 0.95～1.15 and 1.30～1.45 for low and high solar altitude angles, respectively. The cloud-sky segmentation threshold decreases as visibility decreases; when visibility is >10 km and <1 km, the solar altitude angle threshold ranges are 1.07～1.45 and 0.95～1.15, respectively, while the threshold segmentation method loses reliability when visibility is <500m. Validation through interactive scenario threshold verification and cloud type differentiation verification demonstrates that this identification method reduces the mean absolute error of cloud cover identification to 11.0% in complex scenarios with low solar altitude (<10°) and low visibility (<1 km), representing a 36.7% improvement over the fixed threshold method. The recognition accuracies for cloud cover fractions of 0～1, 1～3, 3～7, and >7 tenths reach 93%, 89.5%, 88.7%, and 95.9%, respectively, which are 3.7%, 4.2%, 2.2%, and 2.5% higher than the existing optimal baseline (DeepLab V3+).