Timely and accurate detection of oil spills from Synthetic Aperture Radar (SAR) imagery is essential for marine environmental monitoring and emergency response. This study presents a multi-task deep learning framework that simultaneously performs oil spill segmentation, classification, and ecological risk assessment from Sentinel-1 SAR data. The architecture employs a shared ResNet50 encoder enhanced with Convolutional Block Attention Modules (CBAM) to suppress background clutter and focus on oil-relevant features. A U-Net++ decoder generates pixel-wise segmentation masks, while a Vision Transformer head performs binary oil/non-oil classification. To address multi-class segmentation challenges, a composite loss function combining cross-entropy, Jaccard, Tversky, and boundary losses is introduced. The model is pre-trained on 5,630 Sentinel-1 patches to bridge the domain gap between ImageNet weights and SAR data, followed by fine-tuning on the DARTIS-2019 dataset. Evaluation yields a classification accuracy of 89.06% with corresponding precision of 89.02%, recall of 89.06%, and F1-score of 88.99%, alongside a segmentation mean Intersection-over-Union (mIoU) of 0.38 ± 0.08. While the segmentation performance is comparatively modest; reflecting the inherent challenges of speckle noise, class imbalance, and fine-grained boundary delineation in SAR imagery; it provides sufficient spatial localisation to enable downstream risk assessment. The ecological risk module integrates segmentation outputs with species sensitivity indices and habitat vulnerability data through a weighted scoring function (area, species presence, habitat type, and persistence), producing risk scores categorised as Low, Medium, High, or Critical. Although preliminary, the risk outputs are derived from established ecological sensitivity frameworks and are intended as decision-support indicators rather than definitive impact assessments. Model interpretability is enhanced through Grad-CAM visualisations, providing spatial heatmaps that highlight influential regions and build operator trust. The proposed system offers an integrated, end-to-end solution for oil spill monitoring that unifies detection, classification, and ecological risk evaluation, representing a meaningful advancement in AI-assisted marine conservation and operational remote sensing.