Underwater Salient Object Detection (USOD) has emerged as a specialized division of Salient Object Detection (SOD) to meet the increasing demand for underwater exploration. However, two major challenges hinder its development. First, current USOD datasets are limited in size and diversity, making them insufficient for robust model training and evaluation. Second, most existing deep learning-based USOD approaches rely on convolutional encoder-decoder architectures, which improve inference over traditional methods but still suffer from limited receptive fields and high computational costs. To address these challenges, we introduce a new USOD method based on the Swin transformer, which incorporates a multimodal feature fusion module to increase detection accuracy. Our model uses RGB images and depth maps to capture both semantic information and visual details. The multimodal feature fusion module dynamically adjusts feature importance through attention mechanisms, enhancing the ability of the model to extract meaningful information. By using a Swin transformer model with pretrained weights as the encoder backbone, the proposed method mitigates data scarcity issues and captures both the global context and long-range dependencies. The experimental results demonstrate that our approach achieves significant improvements in detecting salient objects in underwater environments.