The semantic segmentation approach is essential in automated scene analysis, but its application in underwater environments is still limited. Datasets generally have insufficient labeled data, unbalanced data classes, and different lighting conditions, making it difficult to obtain optimal results. Currently, deep convolutional neural networks allow very good results in machine vision tasks, and one of the network architectures with good performance in semantic segmentation is DeepLabv3 +. This paper evaluates the performance of DeepLabv3 + and transfer learning based on pre-trained backend networks in ImageNet to study underwater scenes. The experimentation is carried out on a dataset available on the Internet with labels of eight classes. Experimental results show that DeepLabv3 + and transfer learning are effective for semantic segmentation of multiple underwater scene objects with insufficient tagged data and unbalanced classes.