Cloud segmentation is a prerequisite for reliable analysis and cloud removal workflows in optical remote sensing imagery. In this study, we investigated a lightweight U-Net for binary cloud mask generation on the 38-Cloud dataset utilizing an EfficientNet-B0 encoder adapted for dense prediction. The network was trained using the AdamW optimizer and binary cross-entropy under a short baseline training schedule, resulting in an F1-score of 0.9517, precision of 0.9467, recall of 0.9567, IoU of 0.9079, and accuracy of 0.9698, while using 10.6 million parameters. Compared with heavier segmentation models, the proposed architecture offers a favorable accuracy-to-complexity tradeoff and remains practical for resource-constrained remote sensing workflows. Our findings demonstrate that a carefully optimized encoder–decoder design can deliver competitive cloud segmentation performance without relying on computationally expensive attention or transformer modules.