Application of dimensionality reduction in real-time inundation area prediction using deep learning

Abstract

This paper proposes an alternative model for inundation area prediction using conventional numerical methods by applying dimensionality reduction techniques to inundation area prediction using deep learning. For dimensionality reduction methods、 used singular value decomposition (SVD), non-negative matrix factorization (NMF), and auto encoder (AE) are employed. The proposed method was applied to a simulation of Arakawa River inundation prediction, and its validity and effectiveness were examined from the viewpoints of computational time and accuracy. As a result, it was confirmed that the proposed method can significantly reduce computation time and maintain accuracy. We also confirmed that SVD is the most effective of the three dimensional compression methods employed.