Abrasive belt grinding is widely applied in the industry field. Due to its flexible contact with the workpiece and the nonuniformity of abrasive distribution on the belt, the material removal rate is difficult to be predicted accurately in theory. It directly affects the efficiency and quality control of the abrasive belt grinding. This study proposes a method for identifying the material removal rate of abrasive belt grinding based on spark images. A segmentation algorithm for spark images is presented. Quantitative feature models of the color, brightness, area, and contour features of the spark image are formulated. Pearson coefficient is used to analyze the correlation between the feature of the spark image and the material removal rate of abrasive belt grinding. A linear regression model based on the single feature of the spark image and a multifeature regression model based on the support vector regression (SVR) are established, respectively. The maximum error, the mean square error, and the determination coefficient are used as the evaluation metrics. Experimental results show that the multifeature SVR model based on the radial basis kernel function can achieve high prediction accuracy with the determination coefficient of 0976. The proposed method in this paper provides a new way to effectively control the material removal rate of abrasive belt grinding.