Abstract: Typhoon disaster is one of the most serious marine disasters in China. Research on improving the accuracy of typhoon disaster loss prediction is of great significance to disaster prevention and mitigation.Aiming at the problem that the current machine learning algorithms have low prediction accuracy when processing small sample data, BP neural network algorithm combined with fuzzy mathematics is proposed to predict the loss of typhoon disasters. In this paper, 25 typhoon sample data with better records were selected for experiments in Guangdong province from 2005 to 2016. First, the initial data is diffused with normal information using the information diffusion technology, and then combined with BP neural network to predict the typhoon disaster loss.The results show that this method can solve the problem of fewer measured samples and contradictory samples of typhoon disaster and improve the accuracy of typhoon disaster loss prediction.