The burn index, based on remote sensing images, is widely used to study fire severity. The quantitative evaluation of grassland fire severity by selecting a suitable burn index has great significance when studying vegetation recovery and management. A burnt area of Hulunbuir pasture land was selected as the study area. Landsat8 OLI images were used to develop a regression model of the Composite Burn Index (CBI) using four different burn indices (Normalized Burn Ratio, NBR; NIR-SWIR-Temperature Version1, NSTV1; differenced Normalized Burn Ratio, dNBR; and Relative differenced Normalized Burn Ratio, RdNBR), and the accuracy of the model was tested. The ability to detect fire severity at different levels by the four different burn indices was compared. The results showed that dNBR (n=70, R²=0.856) had the best fitting effect in the polynomial regression model; but there were some differences between the four burn indices when identifying grassland fire intensity at different levels. The NSTV1 index performed best at moderate severity (1 < CBI ≤ 2). For non-fire (CBI=0), low severity (0 < CBI ≤ 1), and high severity (2 < CBI ≤ 3), the dNBR index performed best with accuracies up to 80%, 62.5%, and 100%, respectively; A map of grassland fire severity, which was drawn using the dNBR index, had the highest overall accuracy at up to 82.1%. The Kappa coefficient was also up to 0.76. In conclusion, the dNBR index is the most suitable remote sensing index for analyzing and evaluating grassland fire severity.