Due to the frequent fluctuation of the ore grade and other random disturbances in the flotation process, reagents need to be changed in real time to achieve desired metallurgical parameters. To avoid erratic operation, a data-driven reagent evaluation model was proposed. It is insensitive for bubble size distribution to differentiate froth surface images with a similar average bubble size but different grade. Since bubble shape is heavily influenced by froth grade, and bubble deformation is positively correlated to the froth grade in industry, a novel froth image feature named the union distribution of bubble size and shape was developed. Then, a multi-output least square support vector regressor was introduced to develop a dynamic causal model to simulate the relationship between the reagents and the produced future froth surface appearance feature. Next, the health status of the reagents is rated by recognizing the category of predicted future froth. Under guidance from this reagent evaluation model, the rationality of reagent dosages can be obtained in priority. The proposed method was applied in a gold-antimony flotation plant located in Hunan, China. It improved the efficiency of froth flotation and reduced false reagent operation.