In the digital Polymerase Chain Reaction(dPCR) detection process, discriminating positive droplets from negative ones directly affect the final concentration, which is one of the important factors affecting the accuracy of the instrument. Current methods do not take into account the influence of sample concentration on the result error, resulting in a larger deviation from the actual results at a low concentration. In this paper, a florescent droplets classification method was designed based on generalized Pareto distribution. It was discussed that the possible effects of misclassification at different concentrations on the results, determined the high quantiles of generalized Pareto distribution, and verified the proposed method on the self-made droplet digital PCR. Experimental results showed that for the method proposed, the linear regression of samples with target copies from 5 to 5 000 got an r2=0.995 6 and a detection limit of less than 5 copies/samples, while that of the comparison method was less than 50 copies/sample. These results indicate that the proposed method improves the lower detection limit of the droplet digital PCR by oneorder, and can achieve automated droplet classification at ultra-low concentration.