Abstract: Cloud cover tends to hinder information extraction from remote sensing images during image processing. However, complex and changeable surface backgrounds make it difficult to effectively extract the differences in features between cloud targets and backgrounds. Although existing methods exhibit satisfactory cloud detection effects under most backgrounds, they show significant misclassification and omission in some environments, failing to maintain encouraging performance due to poor stability and insufficient generalization ability. Given this, this study proposed a cloud detection method for multiple backgrounds. Based on Sentinel-2A data, this study analyzed the differences in spectral characteristics between cloud targets and backgrounds to assist in the selection of samples for detection. Based on this, this study introduced more effective detection indices HOT and CDI. Finally, this study obtained a random forest-based cloud detection model through training. Then, from the perspective of the influence of backgrounds and cloud target types on detection accuracy, this study compared the obtained cloud detection model with the Fmask algorithm using images with different backgrounds. The comparison results show that the method proposed in this study increased the overall accuracy and F1 score by 2.2% and 2.9%, respectively, with the standard deviations of them reducing by 29.6% and 72.5%, respectively. These findings indicated that this method can significantly improve the stability of cloud detection in different environments while maintaining high detection accuracy. Therefore, this method is effective in cloud detection in multi-backgrounds.