Breakup of liquid jets in crossflow contain unique embedded patterns based on the type of the pertained flow regime. Recognition of these patterns and correlating them to the underlying flow schemes is a possible but challenging task due to their complex nature. In this research, we have utilized unsupervised reduced-order models to create a set of modes that could be employed to analyze the attributes of different snapshots. They may be imported to feature-based supervised classifier to diagnose multiple flow regimes. These models include proper orthogonal decomposition, principal component analysis, and dynamic mode decomposition (DMD). Snapshots are being extracted by high-speed imaging of the flow field of 14 different cases at various categories. These images are then stacked into a high-dimensional matrix as the training set for the support vector machine and random forest (RF) classifiers to learn. Then, the generated classifiers in the previous step are used to predict which category belongs to every dataset of the six newly imported cases. Afterward, the accuracy level of different permutations of reduced-order models and machine learning algorithms is calculated. Results indicate that using dynamic modes of DMD in partnership with the RF algorithm outperforms every other model with the highest accuracy rate of 95%. Finally, a decision-maker application that classifies the datasets based on the first three models with the highest accuracy levels is introduced to provide a user-friendly environment for data classification at all other potential conditions.