UPE-Net: A Spatial Information Enhanced Network for Ulva Prolifera Extraction From GF-1 WFV Images

The outbreak of Ulva prolifera in the Yellow Sea of China has caused serious economic losses. Strengthening Ulva prolifera disaster monitoring can reduce economic losses, support sustainable marine resource use, and maintain ecological balance. Deep learning-based Ulva prolifera disaster monitoring has gradually become a research hotspot. However, existing deep-learning semantic segmentation methods are insufficient in perceiving the spatial information of Ulva prolifera. On the one hand, downsampling will lose the spatial information of Ulva prolifera, and on the other hand, existing models are difficult to effectively capture the multiscale spatial characteristics of Ulva prolifera. In order to solve these problems, this article proposes an Ulva prolifera extraction network (UPE-Net) based on the U-Net model with three important modifications. First, we develop a novel multifeature downsampling module, MH-downsampling block, specifically designed to mitigate spatial information loss. Second, we propose a dual attention spatial pyramid pooling module that effectively suppresses noise interference while capturing multiscale spatial information on Ulva prolifera. Third, a detail-enhanced attention block is incorporated to enhance the learning of crucial spatial and detail features during the decoding process. The performance of the proposed model was evaluated using the Gaofen-1 wide field view Ulva prolifera dataset. Compared with existing methods, our UPE-Net has significantly enhanced the ability to perceive the spatial information of Ulva prolifera, improving the IoU by at least 8%, the segmentation results of our model are closest to the ground truth. Thereby our UPE-Net can provide a robust basis for monitoring Ulva prolifera disasters.