Many recently proposed robust two-dimensional principal component analysis (2DPCA) approaches can suppress the sensitivity to outliers in images to some extent. However, most approaches can neither perfectly minimize the reconstruction error nor use fewer coefficients to conveniently represent image information. To alleviate these deficiencies, we developed a novel robust 2DPCA approach for underwater image analysis, called l2,p-sequential bilateral-2DPCA (l2,p-SB-2DPCA). The outstanding advantages of l2,p-SB-2DPCA are as follows. First, our model uses the l2,p-norm as the metric criterion of the objective function, which not only improves the robustness of the algorithm but also preserves the basic properties of 2DPCA. Second, we establish the relationship between the variance of the projection data and the corresponding input data in both the row and column directions, which makes the model achieve good recognition ability while using fewer coefficients and further improves the interpretability of the model. Finally, to obtain the optimal value of l2,p-SB-2DPCA, we present an iterative algorithm. The experimental results show that the proposed algorithm achieves the best performance on three underwater datasets and one extended face dataset compared with other robust 2DPCA approaches.