海洋油污染對海岸生態、漁業與沿海經濟造成嚴重且難以回復的傷害，因此能否 快速掌握油污的汙染範圍，攸關後續清污作業的成效。在各類遙測手段中，合成孔徑 雷達（Synthetic Aperture Radar， SAR）因具備全天候、晝夜皆可運作且單幅影像即可涵 蓋廣大海域的特性，被視為最適合油污監測的感測器，其中又以Sentinel-1所提供的免 費開放影像為主要資料來源。隨著影像資料量隨年增加，研究重心轉向以深度學習進 行油污語意分割的自動化辨識。然而，真實觀測場景涵蓋極為廣大的海域，多數公開 資料集為預先裁切成固定尺寸的影像，無法反映真實大尺度影像的複雜度，使現有方 法於實際應用時的分割表現受到限制，無法準確評估。 除此之外，將深度學習方法應用於大尺度SAR油污分割時仍面臨數種困難：像素 標註成本高昂；低風速區與生物膜等海面疑似目標（look-alikes）在外觀上與油污非常 相似，難以辨別；而大尺度影像若直接降採樣將模糊油污絲的細長邊界，若改採局部 切塊雖能保留細節卻喪失全局語意，形成粗粒度與細粒度特徵難以兼顧的問題。為克 服上述挑戰，本研究提出一套針對整個SAR影像的雙階段油污分割框架（ADual-Stage Framework for Oil-Spill Segmentation on Large-Scale SAR Imagery），我們在資料前處理階 段同時保留空間細節並生成具高度判別力的特徵，使模型能夠區分真實油污與疑似目 標。 我們將分為資料前處理與模型訓練兩個階段。在第一階段，改良物理導向三通道 編碼（pseudo-RGB），並引入OpenStreetMap（OSM）陸地遮罩以強化油污邊界對比並 抑制疑似目標；接著對全幅影像進行重疊切塊，並結合ErrorL2-Norm（EL2N）困難度 評分與分層多樣性選擇之資料篩選策略，在沒有標註相似物的情況下，使分割模型聚 焦於易混淆的困難負樣本而不被冗餘背景主導。在第二階段，將篩選後之資料用於分 割模型之訓練與推論，完成大尺度影像的像素級油污分割。 在實驗中，本研究在兩個資料集上進行評估，並與DeepLabV3+、SegFormer與 OSDMamba 等代表性分割方法比較。從自己建立的大尺度衛星雷達油污資料集LS i SAR-Oil 上，我們的方法在所有評估指標皆取得最佳結果，油污IntersectionoverUnion （IoU）達45.8、mean Intersection over Union （mIoU）達 72.9，優於各對照方法，突顯 出在真實大尺度全幅影像上區分高度相似偽影與細粒度油污的能力。然後在公開的固 定裁切的M4D影像資料集上，即使僅使用壓縮後的區塊影像訓練，本方法在二元資料 集限制下，在未標註相似物的情況下油污IoU仍可維持60.4，證實本方法仍對於固定 裁切的樣本具有良好的判釋能力。 最後，我們也透過消融實驗檢測各方法的貢獻。結果顯示，物理導向三通道 編碼相較於原始Vertical-Vertical （VV）單波段可使 mIoU 提升約 5%、95th Percentile Hausdorff Distance （HD95）降低約 20%；在此基礎上再結合資料篩選後，相較於隨機 採樣負樣本，油污IoU進一步提升約5.2%。實驗證實，兼顧邊界對比強化與困難樣本 平衡之前處理設計，能有效改善大尺度SAR影像之油污分割能力，展現本方法於真實 海域監測場景中的實用性與發展潛力。 關鍵字：合成孔徑雷達影像、分割模型、雙階段框架、三通道編碼、資料篩選;Marine oil pollution causes severe and hardly reversible damage to coastal ecosystems， fisheries， and coastal economies; therefore， the ability to rapidly determine the extent of con tamination is crucial to the effectiveness of subsequent cleanup operations. Among the various remote sensing techniques， Synthetic Aperture Radar (SAR) is regarded as the most suitable sensor for oil-spill monitoring， owing to its all-weather， day-and-night operability and the wide sea-surface coverage of a single image. In particular， the free and openly accessible imagery provided by Sentinel-1 has become the primary data source. As the volume of imagery grows year by year， research focus has shifted toward deep-learning-based semantic segmentation for oil-spill recognition. However， real observation scenes cover extremely wide sea areas， whereas most public datasets consist of images that have been pre-cropped to a fixed size and thus fail to reflect the complexity of genuine large-scale imagery， limiting the segmentation performance of existing methods in practical applications. Moreover， applying deep-learning methods to oil-spill segmentation on large-scale SAR imagery still faces several difficulties: pixel-wise annotation is costly; sea-surface look-alikes such as low-wind-speed regions andbiogenic films are visually very similar to oil spills and diffi cult to distinguish; and directly down-sampling a large-scale image blurs the slender boundaries of oil filaments， while local patching preserves details but loses global semantics， creating a dilemma in which coarse-grained and fine-grained features are hard to reconcile. To overcome these challenges， this study proposes A Dual-Stage Framework for Oil-Spill Segmentation on Large-Scale SAR Imagery， in which the data-preprocessing stage simultaneously preserves spa tial detail and generates highly discriminativefeatures， enablingthemodeltodistinguishgenuine oil spills from look-alikes. The proposed method is divided into two stages: data preprocessing and model training. iii In the first stage， we employ a physics-guided three-channel (pseudo-RGB) encoding and in corporate an OpenStreetMap (OSM) land mask to strengthen oil-spill boundary contrast and suppress look-alikes; the full-scene image is then partitioned into overlapping patches， and a data-selection strategy combining EL2N difficulty scoring with stratified diversity selection is applied so that， even without look-alike annotations， the segmentation model focuses on confus able hard negatives rather than being dominated by redundant background. In the second stage， the selected data are used for the training and inference of the segmentation model， completing pixel-wise oil-spill segmentation on large-scale imagery. In the experiments， this study evaluates the method on two datasets and compares it with representative segmentation methods including DeepLabV3+， SegFormer， and OSDMamba. On LS-SAR-Oil， the large-scale satellite SAR oil-spill dataset constructed in this study， our method achieves the best results across all evaluation metrics， reaching an oil-spill IoU of 45.8 and an mIoU of 72.9， significantly outperforming all baselines and highlighting its ability to distinguish highly similar artifacts from fine-grained oil spills on genuine large-scale full-scene imagery. On the publicly available， fixed-crop M4D image dataset， even when trained only on compressed patch images and under a binary-dataset constraint without look-alike annotations， our method still maintains an oil-spill IoU of 60.4， confirming that it retains strong discrimina tion capability on fixed-size samples. Finally， we examinethecontributionofeachcomponentthroughablationexperiments. The results show that， compared with the original single-band VV input， the physics-guided three channel encoding improves mIoU by about 5% and reduces HD95byabout20%; building upon this， adding data selection further improves the oil-spill IoU by about 5.2% relative to randomly sampled negatives. The experiments confirm that a preprocessing design balancing boundary contrast enhancement and hard-sample balancing effectively improves oil-spill segmentation on large-scale SAR imagery， demonstrating the practicality and development potential of the proposed method in real-world sea-area monitoring scenarios. Keywords: Synthetic aperture radar imagery， Semantic segmentation， Dual-stage framework， Three-channel encoding， Data selection
