A Deep Learning Framework for Automatic Detection and Segmentation of Internal Solitary Waves in Satellite Imagery over the Northern South China Sea

Internal solitary waves (ISWs) play an important role in ocean energy transport and mixing, yet their elongated, curved crest lines and low contrast make automated extraction from satellite imagery challenging. To address these challenges, this study proposes YOLO11-ISW, an improved YOLO11-based framework for joint ISW detection and pixel-level segmentation. Dynamic Snake Convolution (DSConv) is introduced to enhance the representation of curved crest-line structures, while the Convolutional Block Attention Module (CBAM) strengthens responses to low-contrast ISW features. A MODIS-based ISW dataset from the northern South China Sea was constructed for model training and evaluation. YOLO11-ISW achieved detection Precision, Recall, and F1-score values of 92.01%, 80.73%, and 86.00%, respectively, and segmentation values of 93.14%, 82.60%, and 87.55%, respectively. The model achieved the highest detection and segmentation F1-scores among the evaluated methods. The segmentation results were further georeferenced to extract ISW crest-line locations, and the framework was applied to MODIS imagery from 2011 to 2020 to investigate satellite-observed ISW spatiotemporal variability. The detected ISWs exhibited pronounced spatial clustering, strong spring–summer seasonality, and clear lunar-cycle modulation, consistent with previous observations in the northern South China Sea. These results demonstrate the potential of YOLO11-ISW for automated, long-term, and large-scale monitoring of ISWs using satellite remote sensing imagery.