Abstract: Water is highly critical to the growth of crops. Water deficit can cause to be lower than normal levels for the water potential and turgor pressure in the crops. The normal metabolic functions can be destroyed in this case, even to seriously threaten the growth and development of crops. Therefore, it is very necessary to timely identify crop drought stress for the better growth of plants, rational irrigation, and yield. Alternatively, chlorophyll fluorescence imaging technology has widely been used to represent the crop photosynthesis data, such as the absorption and transformation of light energy by leaves, the energy transmission and distribution, and the state of the reaction center. Furthermore, various stress states of plants can be early monitored before the symptoms are visible to the naked eyes. Many efforts have been made to identify crop abiotic stress using the chlorophyll fluorescence technology. However, there are still the following challenges: 1) Most studies focused only on the minimum fluorescence and the maximum light quantum efficiency after dark adaptation, but failed to use all chlorophyll fluorescence parameters; 2) Fluorescence parameters were collected at only a few points to assume as the image dataset of one leaf, much less to the entire plant. Therefore, this research aims to systematically identify tomato seedling under different drought levels using chlorophyll fluorescence imaging and machine learning. Firstly, four levels of drought stress were set in the soil moisture content, including the suitable water, mild, moderate, and severe drought. Secondly, the chlorophyll fluorescence imaging system was used to collect the dataset of plant canopy under various drought stress levels. The image pixels were averaged as the chlorophyll fluorescence parameter value of the plant. Successive Projections Algorithm (SPA), Iterative Retained Information Variable (IRIV), and Variable Iterative Space Shrinkage Approach (VISSA) were then used to extract the chlorophyll fluorescence parameters highly related to drought stress. Five common parameters were achieved to analyze the correlation with the drought stress, including the actual light quantum efficiency at L2 time, and the non-actinic fluorescence quenching at L3 time during the light adaptation, the light adaptation-photochemical quenching at L2 time, steady-state light adaptation-photochemical quenching, and the light adaptation-photochemical quenching at D3 time during the dark relaxation in the "Lake" model. Finally, a recognition model of drought stress state was established using Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and k-Nearest Neighbor (KNN). Subsequently, a confusion matrix was also constructed to determine the accuracy of the model for different drought states. The results showed that the LDA recognition model of drought stress presented the highest average recognition accuracy, followed by SVM, and KNN. The modeling accuracy of the selected parameters of SPA, IRIV, and VISSA was equal to or slightly higher than that of the full-parameter modeling. It showed that the selected parameters contain most of the photosynthesis information of plants under drought stress and proves the effectiveness of the fluorescence parameters extracted by the three parameter optimization algorithms. In the LDA drought identification model, the accuracy of IRIV-LDA for identifying suitable moisture, mild drought, moderate drought and severe drought was improved by 6%, 4%, 2% and 2% respectively compared with full parameter-LDA, and the accuracy reached 100%, 95%, 98% and 98% respectively. Consequently, the kinetic parameters of chlorophyll fluorescence can be used to accurately detect the drought stress state of tomato seedlings. This finding can provide a new insight for early monitoring of drought stress and determination of damage levels in tomato seedlings and similar crops.