Detection of offshore oil slicks utilizing a novel joint feature extraction technique that integrates limited hyperspectral image data

Abstract The marine environment is facing severe threats from oil spills, and hyperspectral remote sensing image technology holds great potential for rapid detection of such incidents. However, the limited sample sizes and insufficient extraction of feature information currently hinder accurate identification of offshore oil spills. To address this issue, we present the first implementation of a meta‐learning model in marine remote sensing observations using the meta‐learning–based perceptual attention generative adversarial network model (meta‐PAGAN) method. This approach integrates a perceptual attention‐based generation adversarial network to effectively detect offshore oil films despite the scarcity of hyperspectral sample data. By simultaneously executing the meta‐learning model and employing the perceptual attention network mechanism, reliance on sample data is reduced while overall generalization capability is enhanced, thereby significantly improving detection accuracy. We conducted experiments using airborne visible light/infrared imaging spectrometer (AVIRIS) hyperspectral data collected during the Deepwater Horizon incident on 9 July 2010 to evaluate our meta‐PAGAN model's performance in detecting offshore oil spills with limited hyperspectral data samples. The experimental results demonstrate that our model outperforms other existing models with an impressive identification accuracy rate of 98.4%. Our proposed approach possesses substantial application value in marine environment observation and navigation safety protection.