本研究於 DeepLabv3、MariNeXt 與 OSDMamba 三種不同架構的模型上、以 5 折交叉驗證進行驗證，且訓練樣本採樣與推論評估皆在完整大範圍場景上進行。實驗結果顯示，相較於僅在油污周邊區域採樣的 baseline，本框架在三個模型上的油污交聯比（Intersection over Union， IoU）均顯著提升（如 DeepLabv3 由 0.055 提升至 0.213，三個模型約為三至五倍），平均交聯比（mean IoU， mIoU）、Precision 與 F1 亦同步顯著上升，且此一效益跨三種架構一致；其代價為 Recall 略有下降，反映本框架以較低的 Recall 換取大幅提升的 Precision，以符合大範圍場景下抑制誤報的需求。消融實驗進一步顯示，重疊圖塊採樣為跨架構最穩定的貢獻來源，困難負樣本探勘的效益主要展現於探勘所用之模型，光譜指數過濾則作為標籤雜訊的安全機制。這些結果顯示，在不更動模型架構的前提下，僅透過資料層面的取樣與篩選即可顯著提升大範圍場景的油污判釋效能，並降低對人工挑選訓練樣本的依賴。;Marine oil spills cause severe ecological and economic harm. Unlike tracking that follows a reported incident， routine monitoring cannot know in advance when or where a spill will occur and must scan wide sea areas with satellite imagery， making timely， large-scale detection an important task in environmental monitoring.. Sentinel-2 multispectral satellite imagery， with its wide-area， periodically revisited coverage， combined with deep-learning semantic segmentation， has become a viable route for oil spill detection. However， detecting oil spills over full scenes faces two main difficulties: oil pixels constitute an extremely small fraction of an image， causing severe class imbalance; and the sea surface contains many regions visually similar to oil (look-alikes) that readily induce false positives. Existing multispectral methods mostly train and predict on patches manually delineated around oil regions， which fail to reflect the full-scene setting and rely heavily on expert manual selection.
To reduce the reliance on manual training-sample selection， this study proposes a data-centric framework that generates training data from full， large-scale Sentinel-2 scenes without modifying the underlying deep-learning model. We first construct DV5-MS， a full-scene oil-spill dataset derived from real reported events， which serves as the raw input to the framework; the framework then automatically samples training patches from it through three parts: overlapping patch sampling， which densely samples oil regions to preserve boundary context; Hard Negative Mining (HNM)， which uses a probe model to mine background negatives readily misclassified as oil， replacing the previous manual sample-selection process; and a spectral-index filter， which takes the spectral indices of these negatives as features， measures their similarity to the oil spectral distribution via the Mahalanobis distance， and removes those most similar to oil (suspected mislabeled oil)， serving as a safeguard against label noise injected into training.
We evaluate the framework on three models of different architectures---DeepLabv3， MariNeXt， and OSDMamba---under 5-fold cross-validation， with both training-sample sampling and inference performed on full， large-scale scenes. Compared with a baseline that samples only around oil regions， the framework significantly improves the oil-class Intersection over Union (IoU) on all three models (e.g.， DeepLabv3 from 0.055 to 0.213; roughly three to five times across the three models)， while the mean IoU (mIoU)， Precision， and F1 also rise significantly and consistently across architectures. The cost is a slight decrease in Recall， reflecting a deliberate trade-off of Recall for substantially higher Precision， which is better suited to suppressing false positives in full-scene monitoring. Ablation studies further show that overlapping sampling is the most consistent contributor across architectures， the benefit of HNM concentrates on the probe model， and the spectral-index filter functions as a safeguard against label noise. These results show that， without altering the model architecture， data-level sampling and filtering alone can substantially improve oil spill detection over large-scale scenes while reducing the reliance on manual training-sample selection.
