FSPN: End-to-end full-space pooling weakly supervised network for benthic habitat mapping using remote sensing images

•Weakly supervised learning is introduced into coral reef benthic habitat mapping.•A full-space pooling network for benthic habitat mapping is proposed.•Dense label simulation function can promote high-quality benthic habitat mapping.•A new benthic habitat mapping dataset based on scribble is constructed.•Provides a new paradigm for ecological surveys of coral reefs. Space remote sensing is powerful for conducting large-scale coral reefs benthic habitat mapping (CRBHM). However, it is difficult because of the large spectral differences in remote sensing images of benthic habitats due to the environment, observation time, atmospheric effects, and water column characteristics. While deep learning convolutional neural networks have shown significant advantages in remote sensing image segmentation, the lack of high-precision public datasets and proprietary networks has led to slow progress in deep learning for BHM. Considering that building a densely annotated coral reef BHM dataset would be expensive and difficult to guarantee accuracy, we introduce weakly supervised learning (WSL) to coral reef BHM for the first time. Subsequently, based on real coral reef benthic habitat survey data, a high-quality, sparsely annotated BHM remote sensing dataset (NJUReef) in the Spratly Islands was created. Moreover, an end-to-end full-space pooling network (FSPN) is proposed that uses shape-adaptive pooling layers to construct full-spatial pooling that can efficiently exploit sparse labels for remote context modeling. In addition, a dense label simulation loss (DLSLoss) is proposed and used to guide FSPN using soft relations between labeled and unlabeled pixels and soft pseudo-labels constructed based on low-level features. We performed extensive experiments on the NJUReef and an open-source dataset (BIRNM). Experimental results show that FSPN outperformed the state-of-the-art WSL model by 6.70 % mIoU. Even for the fully supervised BHM task, the average mIoU is 2.57 % higher than that of the mainstream network.