基于高光谱成像技术的油膜边界像素级检测与可视化OACHSSCD

Aims:To address the requirements of oil type fine identification and spatial distribution visualization in oil spill emergency monitoring,this paper proposed a pixel-level oil film boundary detection and visualization method based on near-infrared hyperspectral imaging.Methods:A push-broom hyperspectral imaging system was used to acquire the data of oil film samples.The ACNN-AS model,pre-trained with 33 types of pure oil samples,was deployed in a zero-shot manner to achieve pixel-level oil type classification.An oil film mask was generated using Otsu adaptive threshold segmentation;and boundary extraction was optimized by combining confidence threshold filtering with median filtering.A multi-level visualization was designed;and pixel statistics were used to calculate the average confidence,proportion,and estimated area of oil.Results:For simulated oil spill samples,the visualization results distinguished the oil film regions from uncertain boundary regions and intuitively revealed the contact boundaries between different oil types.Pixel statistics showed that the average classification confidences for oil types No.19,No.24,and No.32 were 51.81%,57.10%,and 51.86%,respectively.Conclusions:This study validates the feasibility of directly deploying a pure-oil pre-trained deep learning model to complex oil film scenarios.Through the proposed visualization method,an intuitive correspondence between classification results and spatial distribution is achieved,providing an oil spill detection method that combines identification accuracy with spatial representation.