Submersible oil is a kind of oil spill hidden under the sea surface in a suspended state. It has poisoned and eroded the marine ecological environment for a long time. However, effective monitoring means and treatment methods have not been formed for submersible oil pollution, which makes its pollution more sudden and harmful than a sea oil spill. Therefore, it is of great significance to studying effective submersible oil identification methods to protect the marine ecological environment. The TSFS in three-dimensional fluorescence spectroscopy has the advantages of no Rayleigh scattering interference and less redundant data in detecting and identifying oil pollutants. The application of the multidimensional correction analysis method to TSFS data is limited because it does not have a trilinear structure. Thus, a new identification method for seawater submersible oil samples was proposed by combining TSFS with a high-order tensor feature extraction algorithm. First, 90 submersible samples were prepared by using organic dispersants and six different kinds of oil products. Then, the TSFS data of samples were collected using an FS920 fluorescence spectrometer, and the data were preprocessed by standardized. Finally, the identification models of submersible oil samples were established by 2D-LDA and 2D-PCA in the high-order tensor feature extraction method. The established model was compared with the identification model established by conventional MCR-ALS-LDA and NPLS-DA. The results show that the submersible oil sample identification models established by 2D-LDA and 2D-PCA have robust and reliable performance, and the accuracy, sensitivity and specificity of the identification models were 100%, 100% and 100%, respectively. In addition, the fine spectral features of the TSFS spectral image matrix in space, statistics, and graphics can be directly extracted by 2D-LDA and 2D-PCA, which brings a more accurate identification basis for distinguishing submersible oil samples. Therefore, compared with the conventional methods based on expansion or decomposition of data, the more accurate prediction results were obtained by the discrimination model established by the high-order tensor feature extraction method. This study provides a reference for submersible oil identification.