Abstract -- The main purpose of this project is to collect different data which result in debris flow, and apply different neural networks to assess its practicability and accuracy that design the debris flow warning system. For the computation of the performance with some different architectures, we attempt to construct a debris-flow warning system using the shared near neighbors (SNN). The SNN can be regard as an unsupervised learning method. The advantage of SNN is that it can deal with the non-globular cluster, in the other words, it means that the data which has non-globular cluster can be partitioned with some specific meanings by its concept of clustering. As the review of past research, we find that there were some specific relations between the occurrence of the debris-flow and precipitation, so we use the characteristic of SNN to match up the hydrology condition of debris flow disaster for simulation some calamity may happening in the future. We will also discuss and improve the problem that the model performance and some partition which the architecture may lack or not consider.