Abstract: In response to the issues of limited training samples and unverified quality of crowdsourced reef and island data when using deep learning models to extract reef and island data from remote sensing imagery, this study proposed a method that combined crowdsourced and remote sensing data for the intelligent extraction of reefs and islands. The research focused on the Maldives and Xisha Islands, where crowdsourced reef and island data was obtained from OpenStreetMap (OSM), and Sentinel-2 imagery was acquired from Google Earth Engine (GEE). The datasets were processed to generate training samples for five models: FCN, U-Net, SENet, HRNet, and SegFormer. The visual interpretation results from the model testing sets were used as ground truth to evaluate the performance of the predictions from five models and the original OSM label data using metrics such as accuracy, precision, recall, and F1 score. Additionally, the intersection-over-union (IoU) was calculated to compare and validate the superiority of the deep learning model predictions. Experimental results demonstrated that the U-Net model achieved F1 scores exceeding 96.00% and 93.00% for islands and reefs in both study areas, yielding superior prediction accuracy compared to other models, thus making the U-Net model suitable for the intelligent extraction of islands and reefs utilizing combined crowdsourced and remote sensing data. In the Maldives study area, the F1 scores of the U-Net model predictions were approximately 6.48% and 4.30%, higher than those of the OSM label data for islands and reefs, respectively. In Xisha Islands studied, the improvements were about 3.24% and 4.24%. These findings indicated that the U-Net model predictions exhibited better data quality than the original OSM label data, effectively correcting misclassifications of certain reef structures as islands and addressing large-scale omissions in the OSM label data.