Abstract: Oil spills on the sea surface and their emulsifications are a serious threat to the marine environment. As a powerful earth observation technology, hyperspectral remote sensing plays a crucial role in the early detection and qualitative analysis of oil spill. This paper addresses the issues of large data volume of hyperspectral data and limited onboard computing resources, and proposes an ocean emulsified oil spill recognition model that combines Incremental Sparse Principal Component Analysis (ISPCA), with the lightweight deep learning network SqueezeNet. The model is suitable for real-time data flow and large-scale datasets, enabling continuous feature extraction and improving response speed. To verify the model’s performance, this study conducted emulsified oil recognition experiments using airborne hyperspectral images obtained from field simulations and real oil spill scenarios. The results showed that the model can effectively identify oil spills and their emulsified substances (oil-in-water, water-in-oil, non-emulsified oil) as well as background seawater in hyperspectral images, with an overall accuracy (Overall Accuracy, OA) of 83.4% and a Kappa coefficient of 0.81. The ISPCA method can process data in batches, avoiding the memory pressure caused by loading all data at once, and shows stronger information retention ability. Through parameter optimization, high recognition accuracy (above 80% for each category) and good spatiotemporal stability can be ensured, with an identification time of only 110 seconds, which is 47% shorter than conventional principal component analysis. In conclusion, this work not only contributes a lightweight oil spill emulsion identification model suitable for large-scale datasets and online learning scenarios, but also provides a new idea for model optimization, which integrating deep learning and real-time environmental factors. It demonstrates the great potential of hyperspectral remote sensing technology in solving disaster emergency problems.