In this paper, we evaluated how different combinations of CNN architectures and encoders perform in the task of cloud segmentation in satellite images. To accomplish that, we selected and fine-tuned four CNN architectures (U-Net, LinkNet, PSPNet, and MA-Net) with six pre-trained encoders (VGG-16, ResNet-50, Inception V4, Densenet-121, MobileNet V2, and EfficientNet B2). We conduct our experiments using the 38-Cloud, a dataset containing 38 Landsat 8 scene images and their ground truths for cloud detection. We carried out the training process until the validation loss stabilized, according to the early stopping criterion, thus providing a comparative analysis of the best models and training strategies to perform cloud segmentation. We evaluated the performance using classic evaluation metrics, i.e., pixel accuracy, IoU, and Dice coefficient. Results showed the evaluated combinations are capable of segmenting clouds with considerable performance. Regardless of the network architecture, the VGG-16 encoder achieved the best results for all considered metrics. Using VGG-16 combined with the MA-Net achieved the best segmentation results and lower false positives (FP) and false negatives (FN) rates. And, despite being a simpler and older architecture, U-Net combined with the VGG-16 encoder obtains competitive results using fewer parameters than MA-Net.