Marine oil spills and several visually similar oceanic phenomena, such as biogenic slicks, low-wind areas, rain cells, and sea ice, often appear as dark patches in wide-swath synthetic aperture radar (SAR) imagery. This visual similarity among dark patches poses significant challenges for the reliable detection and discrimination of oil spills. To address this issue, we propose the Multi-class Dark Object Detection Network (MDODNet), a dedicated deep learning framework designed to simultaneously identify and classify multiple categories of dark features in wide-swath SAR images. The model is trained and validated using a cross-validated dataset comprising 11,040 annotated dark patches extracted from 2,192 Sentinel-1 VV-polarization SAR scenes. The proposed architecture enhances feature representation through an extended backbone and multi-scale feature fusion, while improving detection robustness via an anchor-based head optimized with the Focal-EIoU loss. Evaluations on an independent test dataset demonstrate strong multi-class detection performance, with average precision (AP) values of 87.04% for biogenic slicks, 93.28% for low-wind areas, 88.82% for oil spills, 90.04% for rain cells, and 89.16% for sea ice. For oil spill scenarios, detected regions are further refined using the UNet++ segmentation model to enable precise boundary delineation and quantitative area estimation. Case studies and cross-platform validation experiments further demonstrate the generalization capability and robustness of MDODNet across diverse SAR data. These results indicate that the proposed framework provides a promising solution for operational, large-scale marine oil spill monitoring.